Reliability and Risk Analysis (What Every Engineer Should Know) [2 ed.] 1032309733, 9781032309736

Citation preview

Reliability and Risk Analysis Completely updated for a new edition, this book introduces reliability and risk analysis for both practicing engineers and engineering students at the undergraduate and graduate levels. Since reliability analysis is a multidisciplinary subject, this book draws together a wide range of topics and presents them in a way that applies to most engineering disciplines. Reliability and Risk Analysis, Second Edition, emphasizes an introduction and explanation of the practical methods used in reliability and risk studies, with a discussion of their uses and limitations. It offers basic and advanced methods in reliability analysis that are commonly used in daily practice and provides methods that address unique topics such as dependent failure analysis, importance analysis, and analysis of repairable systems. This book goes on to present a comprehensive overview of modern probabilistic life assessment methods such as Bayesian estimation, system reliability analysis, and human reliability. End-of-chapter problems and a solutions manual are available to support any course adoptions. This book is refined, simple, and focused on fundamentals. The audience is the beginner with no background in reliability engineering and rudimentary knowledge of probability and statistics. It can be used by new practitioners, undergraduates, and first-year graduate students.

What Every Engineer Should Know Series Editor: Phillip A. Laplante Pennsylvania State University

What Every Engineer Should Know about MATLAB® and Simulink® Adrian B. Biran

Green Entrepreneur Handbook: The Guide to Building and Growing a Green and Clean Business Eric Koester

What Every Engineer Should Know about Cyber Security and Digital Forensics Joanna F. DeFranco

What Every Engineer Should Know about Modeling and Simulation Raymond J. Madachy and Daniel Houston

What Every Engineer Should Know about Excel, Second Edition J.P. Holman and Blake K. Holman

Technical Writing: A Practical Guide for Engineers, Scientists, and Nontechnical Professionals, Second Edition Phillip A. Laplante

What Every Engineer Should Know About the Internet of Things Joanna F. DeFranco and Mohamad Kassab

What Every Engineer Should Know about Software Engineering Phillip A. Laplante and Mohamad Kassab

What Every Engineer Should Know About Cyber Security and Digital Forensics Joanna F. DeFranco and Bob Maley

Ethical Engineering: A Practical Guide with Case Studies Eugene Schlossberger

What Every Engineer Should Know About Data-Driven Analytics Phillip A. Laplante and Satish Mahadevan Srinivasan

Reliability and Risk Analysis Mohammad Modarres and Katrina Groth For more information about this series, please visit: www.routledge.com/What-Every-Engineer-Should-Know/book-series/CRCWEESK

Reliability and Risk Analysis Second Edition

Mohammad Modarres Katrina Groth

MATLAB  is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks does not warrant the accuracy of the text or exercises in this book. This book’s use or discussion of MATLAB  software or related products does not constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular use of the MATLAB  software. Designed cover image: Vincent Paglioni and Katrina Groth Second edition published 2023 by CRC Press 6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742 and by CRC Press 4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN CRC Press is an imprint of Taylor & Francis Group, LLC © 2023 Mohammad Modarres and Katrina Groth First edition published by CRC Press 1992 Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, access www.copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978750-8400. For works that are not available on CCC please contact [email protected] Trademark notice: Product or corporate names may be trademarks or registered trademarks and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging‑in‑Publication Data Names: Modarres, M. (Mohammad), author. | Groth, Katrina, author. Title: What every engineer should know about reliability and risk analysis / Mohammad Modarres, Katrina Groth. Description: Second edition. | Boca Raton : CRC Press, [2023] | Includes index. Identifiers: LCCN 2022050183 (print) | LCCN 2022050184 (ebook) | ISBN 9781032309729 (pbk) | ISBN 9781032309736 (hbk) | ISBN 9781003307495 (ebk) Subjects: LCSH: Reliability (Engineering) | Risk assessment. Classification: LCC TA169 .M63 2023 (print) | LCC TA169 (ebook) | DDC 620/.00452—dc23/eng/20221018 LC record available at https://lccn.loc.gov/2022050183 LC ebook record available at https://lccn.loc.gov/2022050184 ISBN: 978-1-032-30973-6 (hbk) ISBN: 978-1-032-30972-9 (pbk) ISBN: 978-1-003-30749-5 (ebk) DOI: 10.1201/9781003307495 Typeset in Times by codeMantra

Dedication Dedicated to all engineers who understand the importance of the field and continue to advance it in research and practice.

Contents Preface to the Second Edition................................................................................... xv Authors.....................................................................................................................xix Chapter 1 Reliability Engineering Perspective and Fundamentals.......................1 Why Study Reliability?..............................................................1 History and Evolution of the Field.............................................2 Reliability Modeling Approaches...............................................4 1.3.1 Physics of Failure Approach..........................................6 1.3.1.1 Performance-Requirement Model.................6 1.3.1.2 Stress-Strength Model...................................7 1.3.1.3 Damage-Endurance Model............................7 1.3.2 Failure Agents...............................................................7 1.4 Definitions..................................................................................8 1.4.1 Reliability......................................................................9 1.4.2 Risk............................................................................. 10 1.4.3 Availability and Maintainability................................. 11 1.4.4 Failure Modes and Failure Mechanisms..................... 11 1.5 Systems, Functions, and Failure............................................... 12 1.5.1 Failure Modes and Mechanisms for Mechanical Equipment............................................... 15 1.5.2 Failure Modes and Mechanisms for Electrical Equipment................................................... 16 1.5.3 Human Functions, Failure, and Reliability.................20 1.6 Putting It All Together: Risk Assessment Modeling................ 22 1.7 Exercises...................................................................................24 References...........................................................................................24 1.1 1.2 1.3

Chapter 2 Basic Reliability Mathematics: Probability......................................... 27 2.1 2.2 2.3 2.4

2.5

Introduction.............................................................................. 27 Events and Random Variables Used in Reliability................... 27 Sets and Boolean Algebra........................................................28 Probability Terminology and Interpretations........................... 32 2.4.1 Classical Interpretation of Probability (Equally Likely Concept, or Sample Space Partitioning).......... 33 2.4.2 Frequency Interpretation of Probability......................34 2.4.3 Subjective Interpretation of Probability (Bayesian Probability; Evidential Probability)............34 Laws and Mathematics of Probability...................................... 35 2.5.1 Definitions................................................................... 35 2.5.2 Axioms of Probability and Their Implications........... 37 vii

viii

Contents

2.5.3 Mathematics of Probability ........................................ 37 Probability Distribution Basics................................................44 2.6.1 Probability Distributions Defined ..............................44 2.6.2 Distribution Properties ............................................... 49 2.6.2.1 Central Tendency ........................................ 49 2.6.2.2 Dispersion, Shape, and Spread ................... 51 2.6.2.3 Covariance and Correlation ........................ 52 2.6.2.4 Algebra of Expectations ............................. 53 2.7 Probability Distributions Used in Reliability .......................... 53 2.7.1 Discrete Distributions................................................. 53 2.7.1.1 Binomial Distribution ................................. 53 2.7.1.2 Poisson Distribution.................................... 58 2.7.1.3 Geometric Distribution ............................... 61 2.7.2 Continuous Distributions............................................ 63 2.7.2.1 Exponential Distribution............................. 63 2.7.2.2 Continuous Uniform Distribution ............... 65 2.7.2.3 Normal Distribution.................................... 65 2.7.2.4 Lognormal Distribution .............................. 71 2.7.2.5 Weibull Distribution ................................... 74 2.7.2.6 Gamma Distribution (and Chi-Squared) .... 75 2.7.2.7 Beta Distribution......................................... 75 2.7.3 Truncated Distributions ..............................................80 2.7.4 Multivariate Distributions .......................................... 83 2.8 Exercises .................................................................................. 87 References .......................................................................................... 91 2.6

Chapter 3

Elements of Component Reliability ................................................... 93 3.1

Definitions for Reliability ........................................................ 93 3.1.1 Reliability Function .................................................... 93 3.1.2 MTTF, MRL, MTBF, and Quantiles ......................... 95 3.1.3 Hazard Rate and Failure Rate ....................................96 3.2 Common Distributions in Component Reliability................. 101 3.2.1 Exponential Distribution and Poisson Distribution... 101 3.2.2 Weibull Distribution ................................................. 103 3.2.3 Gamma Distribution................................................. 107 3.2.4 Normal Distribution ................................................. 109 3.2.5 Lognormal Distribution............................................ 109 3.2.6 Beta Distribution ...................................................... 110 3.3 Exercises ................................................................................ 111 References ........................................................................................ 114 Chapter 4

Basic Reliability Mathematics: Statistics ......................................... 115 4.1 4.2

Introduction ........................................................................... 115 Descriptive Statistics ............................................................. 116

ix

Contents

4.3 4.4

Empirical Distributions and Histograms................................ 117 Parameter Estimation: Point Estimation................................ 119 4.4.1 Method of Moments.................................................. 120 4.4.2 Linear Regression...................................................... 120 4.4.3 Maximum Likelihood Estimation............................. 120 4.4.4 Bayesian Parameter Estimation................................. 123 4.5 Parameter Estimation: Interval Estimation............................ 124 4.5.1 Confidence Intervals................................................. 124 4.5.2 Credible Intervals...................................................... 127 4.6 Hypothesis Testing and Goodness of Fit................................ 128 4.6.1 Hypothesis Testing Basics......................................... 129 4.6.2 Chi-Squared Test....................................................... 130 4.6.3 Kolmogorov-Smirnov (K-S) Test.............................. 134 4.7 Linear Regression................................................................... 135 4.8 Exercises................................................................................. 140 References......................................................................................... 144 Chapter 5 Reliability Data Analysis and Model Selection................................ 145 5.1

5.2 5.3

5.4

Context and Types of Data..................................................... 145 5.1.1 Types of Field and Test Data..................................... 145 5.1.2 Complete Data........................................................... 146 5.1.3 Censored Data........................................................... 146 5.1.3.1 Left, Right, and Interval Censoring.......... 146 5.1.3.2 Type I Censoring....................................... 147 5.1.3.3 Type II Censoring...................................... 148 Reliability Data Sources......................................................... 148 Nonparametric and Plotting Methods for Reliability Functions................................................................................ 150 5.3.1 Nonparametric Procedures for Reliability Functions................................................................... 150 5.3.1.1 Nonparametric Component Reliability Estimation Using Small Samples.............. 151 5.3.1.2 Nonparametric Component Reliability Estimation Using Large Samples............... 152 5.3.2 Probability Distribution Plotting Using Life Data.... 154 5.3.2.1 Exponential Distribution Probability Plotting....................................................... 155 5.3.2.2 Weibull Distribution Probability Plotting....................................................... 157 5.3.2.3 Normal and Lognormal Distribution Probability Plotting.................................... 158 Maximum Likelihood Estimation of Reliability Distribution Parameters.......................................................... 160 5.4.1 Elements of MLE Using Reliability Data................. 161

x

Contents

5.4.2

Exponential Distribution MLE Point Estimation ..... 165 5.4.2.1 Type I Life Test with Replacement ........... 165 5.4.2.2 Type I Life Test without Replacement ...... 166 5.4.2.3 Type II Life Test with Replacement.......... 166 5.4.2.4 Type II Life Test without Replacement..... 166 5.4.3 Exponential Distribution Interval Estimation .......... 168 5.4.4 Normal Distribution ................................................. 172 5.4.5 Lognormal Distribution............................................ 172 5.4.6 Weibull Distribution ................................................. 174 5.4.7 Binomial Distribution............................................... 177 5.5 Classical Nonparametric Distribution Estimation................. 179 5.5.1 Confidence Intervals for cdf and Reliability Function for Complete and Censored Data .............. 179 5.5.2 Confidence Intervals of Reliability Function for Censored Data .......................................................... 183 5.6 Bayesian Estimation Procedures ........................................... 186 5.6.1 Estimation of the Parameter of Exponential Distribution............................................................... 188 5.6.1.1 Selecting Hyperparameters for the Gamma Prior Distribution ........................ 191 5.6.1.2 Uniform Prior Distribution ....................... 193 5.6.1.3 Jeffrey’s Prior Distribution ....................... 194 5.6.2 Bayesian Estimation of the Parameter of Binomial Distribution............................................... 195 5.6.2.1 Standard Uniform Prior Distribution........ 195 5.6.2.2 Beta Prior Distribution ............................. 197 5.6.2.3 Jeffrey’s Prior Distribution ....................... 199 5.6.2.4 Lognormal Prior Distribution ................... 199 5.6.3 Bayesian Estimation of Other Distributions.............202 5.7 Exercises ................................................................................ 203 References ........................................................................................209 Chapter 6

System Reliability Analysis ............................................................. 213 6.1 6.2

6.3 6.4

Important Notes and Assumptions ........................................ 214 Reliability Block Diagram Method ....................................... 214 6.2.1 Series System ........................................................... 215 6.2.2 Parallel Systems ....................................................... 217 6.2.3 k-Out-of-n Redundant Systems ................................ 220 6.2.4 Standby Systems....................................................... 221 6.2.5 Load-Sharing Systems .............................................224 Complex System Evaluation Methods ................................... 226 6.3.1 Decomposition Method ............................................ 227 6.3.2 Path Tracing Methods (Path Sets, Cut Sets)............. 228 Fault Tree and Success Tree Methods ................................... 232 6.4.1 Fault Tree Method .................................................... 232

Contents

xi

6.4.2 Evaluation of Fault Trees........................................... 236 6.4.2.1 Analysis of Logic Trees Using Boolean Algebra��������������������������������������� 236 6.4.2.2 Combinatorial (Truth Table) Technique for Evaluation of Logic Trees������240 6.4.2.3 Binary Decision Diagrams........................ 242 6.4.3 Success Tree Method.................................................248 6.5 Event Tree Method................................................................. 251 6.5.1 Construction of Event Trees...................................... 251 6.5.2 Evaluation of Event Trees.......................................... 253 6.6 Event Sequence Diagram Method.......................................... 254 6.7 Failure Modes and Effects Analysis....................................... 256 6.7.1 Objectives of FMEA................................................. 256 6.7.2 FMEA/FMECA Procedure....................................... 257 6.7.3 FMEA Implementation............................................. 258 6.7.3.1 FMEA Using MIL-STD-1629A................. 258 6.7.3.2 FMEA Using SAE J1739........................... 261 6.7.4 FMECA Procedure: Criticality Analysis..................264 6.7.4.1 Failure Probability Failure Rate Data Source������������������������������������������������������� 264 6.7.4.2 Failure Effect Probability β.......................266 6.7.4.3 Failure Rate λ p...........................................266 6.7.4.4 Failure Mode Ratio α.................................266 6.7.4.5 Operating Time T......................................266 6.7.4.6 Failure Mode Criticality Number Cm .........266 6.7.4.7 Item Criticality Number ( Cr )..................... 267 6.8 Exercises................................................................................. 269 References......................................................................................... 279 Chapter 7 Reliability and Availability of Repairable Components and Systems.................................................................. 281 7.1 Definition of Repairable System and Types of Repairs�������������������������������������������������������������������� 282 7.2 Variables of Interest: Availability, ROCOF, MTBF............... 282 7.3 Repairable System Data......................................................... 283 7.4 Stochastic Point Processes (Counting Processes)...................284 7.4.1 Homogeneous Poisson Process................................. 287 7.4.2 Renewal Process........................................................ 288 7.4.3 Nonhomogeneous Poisson Process........................... 289 7.4.4 Generalized Renewal Process................................... 291 7.5  Data Analysis for Point Processes............................................ 293 7.5.1  Parameter Estimation for the HPP.............................. 293 7.5.2 Data Analysis for the NHPP..................................... 293 7.5.2.1 Maximum Likelihood Procedures............. 293 7.5.2.2 Laplace Test............................................... 296

xii

Contents

7.6 Availability of Repairable Systems........................................300 7.6.1 Instantaneous (Point) Availability of Revealed-Fault Items�������������������������������������������������301 7.6.2 Instantaneous (Point) Availability of PeriodicallyTested Items�������������������������������������������������������������� 302 7.6.3 Limiting Point Availability.......................................304 7.6.4 Average Availability..................................................304 7.6.5 Other Average Measures of Availability................... 305 7.6.5.1 Inherent Availability.................................. 305 7.6.5.2 Achieved Availability................................306 7.6.5.3 Operational Availability............................306 7.7 Markov Processes for System Availability.............................306 7.8 Availability of Complex Systems........................................... 310 7.9 Exercises................................................................................. 314 References......................................................................................... 319 Chapter 8 Selected Advanced Topics in Reliability and Risk Analysis............ 321 8.1 Uncertainty Analysis.............................................................. 321 8.1.1 Steps in Uncertainty Analysis for Risk and Reliability Models�����������������������������������������������������322 8.1.2 Uncertainty Propagation Methods............................ 326 8.1.2.1 Method of Moments for Propagation of Uncertainties����������������������������������������������326 8.1.2.2 Monte Carlo Methods for Propagation of Uncertainties���������������������330 8.2 Analysis of Dependent Failures.............................................. 335 8.2.1 Single-Parameter Models.......................................... 338 8.2.2 Multiple-Parameter Models.......................................340 8.2.3 Data Analysis for CCFs............................................. 342 8.3 Importance Measures.............................................................344 8.3.1 Birnbaum Importance...............................................344 8.3.2 Criticality Importance...............................................346 8.3.3 Fussell-Vesely Importance........................................ 347 8.3.4 Risk Reduction Worth Importance........................... 350 8.3.5 Risk Achievement Worth Importance....................... 351 8.4 Human Reliability Analysis................................................... 353 8.4.1 The HRA Process...................................................... 353 8.4.2 Identifying Human Failure Events............................ 354 8.4.2.1 THERP...................................................... 355 8.4.2.2 SPAR-H...................................................... 356 8.4.2.3 IDAC and Phoenix..................................... 356 8.4.2.4 IDHEAS.................................................... 359 8.4.3 Representing and Quantifying HFEs........................360 8.4.3.1 THERP...................................................... 362 8.4.3.2 SPAR-H...................................................... 362

Contents

xiii

8.4.3.3 IDAC and Phoenix..................................... 362 8.4.3.4 IDHEAS.................................................... 367 8.5 Exercises................................................................................. 367 References......................................................................................... 373 Chapter 9 Risk Analysis..................................................................................... 377 9.1 QRA Defined.......................................................................... 379 9.2 QRA Process.......................................................................... 382 9.2.1 Initiating Event.......................................................... 383 9.2.2 Hazard Exposure Scenarios...................................... 383 9.2.3 Identification of Hazards........................................... 383 9.2.4 Identification of Barriers........................................... 384 9.2.5 Identification of Challenges to Barriers.................... 384 9.3 PRA Process........................................................................... 384 9.3.1 Objectives and Methodology..................................... 385 9.3.2 Familiarization and Information Assembly.............. 386 9.3.3 Identification of Initiating Events.............................. 387 9.3.4 Scenario Development............................................... 389 9.3.5 Causal Logic Modeling............................................. 390 9.3.6 Failure Data Collection, Analysis, and Assessment....391 9.3.7 Consequence Analysis............................................... 392 9.3.8 Quantification and Integration.................................. 393 9.3.9 Uncertainty Analysis................................................. 395 9.3.10 Sensitivity Analysis................................................... 396 9.3.11 Risk Ranking and Importance Analysis................... 396 9.3.12 Interpretation of Results............................................ 398 9.4 Strengths of PRA.................................................................... 399 9.5 Example: Simple Fire Protection PRA...................................400 9.5.1 Objectives and Methodology.....................................400 9.5.2 System Familiarization and Information Assembly.... 400 9.5.3 Identification of Initiating Events.............................. 401 9.5.4 Scenario Development............................................... 401 9.5.5 Causal Logic Model Development............................402 9.5.6 Failure Data Analysis................................................403 9.5.7 Quantification and Interpretation..............................403 9.5.8 Consequences Evaluation..........................................408 9.5.9 Risk Calculation and Interpretation of Results.........408 9.6 Exercises.................................................................................409 References......................................................................................... 413 Appendix A: Statistical Tables............................................................................ 415 Appendix B: Generic Failure Data...................................................................... 431 Index....................................................................................................................... 453

Preface to the Second Edition Reliability engineering and risk analysis are essential branches of science and engineering that enable us to uphold the integrity of engineered systems and the engineering profession. These capabilities are relevant to various engineering disciplines, including mechanical, chemical, aerospace, nuclear, oil and gas, transportation, electrical, and more. Since engineering designs and processes impose various hazards, risk and reliability analyses are universally important. Interest in reliability and risk analysis has increased due to accidents that have recently resulted in significant public attention, such as the nuclear accidents at Three Mile Island and Fukushima Daiichi, the Deepwater Horizon oil spill in the Gulf of Mexico, the Space Shuttle Challenger disintegration, the Bhopal chemical disaster in India, the Concord crash in Paris, the 2016 and 2021 Tesla’s self-driving cars crashes, and the San Bruno pipeline explosion. The nature of reliability and risk make this an inherently multidisciplinary field because systems and their failures involve the interplay between machines, software, humans, organizations, policy, economics, and the physical world. As systems become more complex and interconnected, learning the methods used to prevent their failures becomes more essential. This book provides a comprehensive introduction to reliability and an overview of risk analysis for engineering students and practicing engineers. The focus is on the fundamental methods commonly used in practice, along with numerous examples. The book’s emphasis is on the introduction and explanation of the practical concepts and techniques used in reliability and risk studies and discussion of their use. These concepts form the foundation of a degree focused on reliability and prepare the student for coursework on advanced reliability methods. This book corresponds to the first course a graduate student takes in reliability engineering and is also suitable for a senior-level undergraduate engineering course or self-study by a practitioner. This book assumes that the readers have taken multivariate calculus and several years of engineering coursework and have some basic familiarity with probability and statistics. The first edition was developed by Mohammad Modarres over 30 years ago from material he developed presented over 10 years in undergraduate and graduate-level courses in Reliability Analysis and Risk Assessment at the University of Maryland. This new version of the book restructures and updates the content based on the further development of the course by Katrina Groth over the past 5 years. Information has been streamlined and simplified to enhance understanding of the foundational elements of the discipline. The notation has been standardized across chapters, and some outdated methods and examples have been removed and replaced with a discussion of more current techniques. Chapters have been condensed and reorganized to enhance readability and understanding. Some aspects of the book have been changed considerably. We structured the book to provide a series of fundamental elements that build upon each other, and thus each chapter builds upon the chapters that come before it. The introduction chapter (Chapter 1) defines reliability, availability, and risk analysis. xv

xvi

Preface to the Second Edition

Chapter 2 introduces the probability mathematics necessary to conduct reliability engineering, while Chapter 3 explains how the basic probabilistic reliability methods characterize the reliability of a basic engineering unit (i.e., a component). Chapter 4 introduces the statistical techniques used to conduct reliability data analysis. In Chapter 5, we discuss the uses of probability, statistics, and data analysis techniques to assess and predict component reliability. In Chapter 6, we present the analytical methods used to assess the reliability of systems (i.e., an engineering unit consisting of many interacting components). The availability and reliability considerations for repairable systems are discussed in Chapter 7. This chapter also explains the corresponding use of the analytical and statistical methods discussed in the earlier chapters and connects them to performing availability analysis of repairable components and engineering systems. In Chapter 8, we introduce four important advanced topics sometimes overlooked in introductory texts but which we believe are foundational to the field and essential for dealing with modern systems. The topics covered are uncertainty analysis, dependent failures, importance measures, and human reliability analysis. While Chapter 8 does not cover the topics in depth, the description should whet the reader’s appetite and motivate further study. This book concludes with Chapter 9, which brings these pieces together under the umbrella of risk analysis, a process that integrates nearly all the analytical methods explained in the preceding eight chapters. For a 16-week undergraduate course, we recommend covering the material in Chapters 1–6 and concluding with Chapter 9. A 16-week graduate course in reliability engineering should be able to cover the entire book. Once these fundamentals are mastered, a student is prepared to take subsequent coursework which delves into more advanced topics in reliability engineering, such as comprehensive probabilistic risk assessment, quantifying and propagating parameter uncertainties, physics of failure, human reliability analysis, dynamic risk assessment, machine learning, prognosis and health management. This book could not have been materialized without the help of numerous colleagues and graduate students since the publication of its first edition. Of particular note are Vincent Paglioni and Yu-Shu Hu, who were instrumental in compiling and proofreading the manuscript and drawing the figures. Numerous other students and colleagues contributed to varying degrees; it is not easy to name all, but we are grateful for their contributions. In addition, we are grateful to V. Paglioni, T. Williams and S. Karunakaran for drawing figures for the second edition. Finally, we are grateful to

Preface to the Second Edition

xvii

our spouses for their endless patience with the long hours and disruption that came with writing this book and for reminding us why it all matters. Mohammad Modarres and Katrina Groth MATLAB® is a registered trademark of The Math Works, Inc. For product ­information, please contact: The Math Works, Inc. 3 Apple Hill Drive Natick, MA 01760-2098 Tel: 508-647-7000 Fax: 508-647-7001 E-mail: [email protected] Web: http://www.mathworks.com

Authors Mohammad Modarres is a scientist and educator in risk and reliability engineering. He is the Nicole Y. Kim Eminent Professor at the University of Maryland. Dr. Modarres co-founded the world’s first degree-granting graduate curriculum in Reliability Engineering. As the University of Maryland Center for Risk and Reliability (CRR) director, Professor Modarres serves as an international expert on reliability and risk analysis. He has authored numerous textbooks, book chapters, and hundreds of scholarly papers. His areas of research interests include probabilistic risk assessment, degradation and physics of failure of materials, nuclear safety, and fracture mechanics. His works in structural integrity and prognosis and health management include both experimental and probabilistic model development. He is a University of Maryland Distinguished Scholar-Teacher, a fellow of the Institute of Electrical and Electronics (IEEE), a fellow of the American Nuclear Society (ANS), a fellow of the Asia-Pacific AI Association (AAIA), and a recipient of the prestigious ANS Tommy Thompson Award for nuclear safety. Dr. Modarres received his BS in Mechanical Engineering from Tehran Polytechnic, MS in Mechanical Engineering, and MS and Ph.D. in Nuclear Engineering from Massachusetts Institute of Technology. Katrina Groth is an associate professor of Mechanical Engineering and the associate director of the Center for Risk and Reliability at the University of Maryland. Her research focuses on engineering safer systems through risk assessment and reliability engineering, with direct impact on hydrogen fueling stations, hydrogen storage, gas pipelines, aviation, and nuclear power plants. She has published over 130 archival papers and reports, authored multiple software packages, and holds 2 patents. Prior to joining UMD, Groth was a principal R&D engineer at Sandia National Laboratories, where she invented the HyRAM (Hydrogen Risk Assessment Models) toolkit used in dozens of countries to establish hydrogen safety standards. Groth holds a B.S. in Engineering and M.S. and Ph.D. in Reliability Engineering from the University of Maryland. She is the recipient of an NSF CAREER award, a technical leadership award from the U.S. Department of Energy’s Hydrogen and Fuel Cell Technology Office, and the 2021 David Okrent Award for Nuclear Safety from the American Nuclear Society. Groth believes that risk assessment should be in every engineer’s toolbox.

xix

1 1.1

Reliability Engineering Perspective and Fundamentals

WHY STUDY RELIABILITY?

Structures, systems, and components designs are not perfect. While a naïve view would insist that it is technologically possible to engineer a perfect system and design out all failures, unfortunately, this view is idealistic, impractical, and economically infeasible. It is an essential job of engineers to prevent, mitigate, respond to, and recover from failures. Reliability engineering helps ensure robust design, product reliability, and economic competitiveness. Beyond this, reliability engineering helps engineers achieve their duty to hold paramount the safety, health, and welfare of the public, workers, and the environment. Reliability engineering and risk analysis are essential capabilities that enable us to uphold the integrity of the systems we design, build, operate, maintain, and ultimately attain the high integrity of the engineering profession. Myriad examples of engineering failures motivate the study of reliability. For instance, on March 11, 2011, an earthquake and tsunami induced a loss of reactor cooling at Japan’s Fukushima Daiichi nuclear power plant. This resulted in meltdown, evacuations, radiation release, and environmental contamination. The cleanup continued a decade later with a cost estimated to exceed $200 billion. Another example comes from the space sector: on February 1, 2003, the Space Shuttle Columbia disintegrated on reentry. The failure scenario was initiated when a small piece of foam insulation broke off during takeoff, causing damage that allowed hot gases to penetrate and destroy the wing. The failure led to the loss of all seven crew members, and the U.S. space shuttle program was halted for over 2 years while NASA investigated the accident and implemented new safety measures. While the Space Shuttle Columbia incident resulted in disaster, NASA was aware that problems with foam had been happening for years. More minor failures happen—and careful study of these is an opportunity to prevent catastrophic failure. An example from the automotive sector demonstrates additional reasons for studying reliability. Widespread Firestone/Ford tire failures in the 1990s led to a recall in 2000. It is estimated that these tire failures cost Firestone’s parent company, Bridgestone, $1.6+ billion and Ford Motor Company $500+ million. In addition, severe financial impacts extended beyond the immediate losses: subsequent corporate restructuring cost Bridgestone $2 billion. The failures also resulted in new regulatory oversight with the passage of the Transportation Recall Enhancement, Accountability and Documentation (TREAD Act), which requires vehicle manufacturers to report safety recalls and defects to the National Highway & Transportation Safety Administration and holds vehicle manufacturers criminally liable if these reporting requirements are violated. DOI: 10.1201/9781003307495-1

1

2

Reliability and Risk Analysis

These are a few examples among many that motivate our studies. Other incidents and near-misses have happened in many industries. In the nuclear industry, the Three Mile Island core damage in 1979 was a well-known disaster and regulatory turning point for the U.S. nuclear industry. For NASA, the 1986 accident of the Space Shuttle Challenger preceded the Columbia disaster. In the chemical, oil, and gas industry, Bhopal, Deepwater Horizon, Texas City, and Piper Alpha accidents have led to significant economic, environmental, and regulatory consequences. Building fires such as the Grenfell Tower fire and the Station nightclub fire led to significant loss of life. On the aviation side: Tenerife, Air France 447, and numerous other crashes have led to the loss of human life and changes to the aviation system. Reliability consideration is not limited only to technologies with catastrophic failure potential. Designers, manufacturers, and end-users strive to minimize the occurrence and recurrence of failures in the components, devices, products, and systems we encounter daily. However, even if technical knowledge is not a limiting factor in designing, manufacturing, constructing, and operating a failure-free design, the cost of development, testing, materials, and engineering analysis may far exceed economic prospects for such a design. Therefore, practical and financial limitations dictate the use of less than perfect designs. A reliable design remains operational and attains its objective without failure during its mission time. However, achieving a reliable design first requires us to understand what an item is intended to do and why and how failures occur. Further, to maximize system performance and efficiently use resources, it is also critical to know how often such failures occur. Finally, it is important to understand various strategies to reduce failures. When failure is understood and appropriately considered in the life cycle of a design, the impact or occurrence rate can be minimized. Further, the system can be protected against failure impacts through careful engineering and economic analysis.

1.2

HISTORY AND EVOLUTION OF THE FIELD

Interest in establishing a quantitative measure for design quality began during World War II, as did the design concept that a chain is only as strong as its weakest link. However, the ideas were primitive and ad hoc. After the war, between 1945 and 1950, the U.S. Air Force became concerned about the quality of electronic products. The starting point of a formal reliability engineering practice is traced back to the Ad Hoc Group on Reliability of Electronic Equipment, established in December 1950. However, the formation and active involvement of the Advisory Group on the Reliability of Electronic Equipment (AGREE) by the U.S. Department of Defense from 1956 to 1958 is considered the turning point in modern reliability engineering. Around 1953, the applications of the probabilistic notions of reliability represented by the exponential distribution became systematically and widely used. One of the main driving forces for this popularity was the simplicity of the corresponding reliability functions. This simplicity accelerated many improvements in traditional statistical and probabilistic approaches to measuring, predicting, and testing item reliability in the 1950s. The reliability block diagram (RBD) concept was adopted from other engineering applications and used to assess system reliability.

Reliability Engineering Perspective and Fundamentals

3

By the 1960s, the exponential distribution proved impractical for many applications and because it is sensitive to departure from the initial assumptions of constant failure rate. Applying this model for components with high reliability targets could result in unrealistic mean time to failure (MTTF) estimates. This was because the exponential model ignores any aging and degradation accumulated in the item. Distributions that considered degradation and wear-out were used, such as the Weibull distribution. The physics of failure (PoF) concept was introduced in the early 1960s as part of a series of symposia at the Rome Air Development Center (which later became the U.S. Air Force’s Rome Laboratory). However, because of its complexity, it was not until the late 1980s that PoF became a serious alternative to statistical reliability prediction methods. The concept of failure modes and effects analysis (FMEA), although introduced by the U.S. military in the late 1940s, was revived by NASA in the early 1960s for its Apollo and other missions as an inductive approach to the analysis of complex systems for which the RBD approach was inadequate. The 1970s experienced more complex system analysis methods, the most important of which was the deductive fault tree analysis approach. This approach was motivated by the aerospace industry’s need for risk and safety assessment and later applied to assess nuclear power plant risks. In this decade, there was intense interest in system-level safety, risk, and reliability in different applications such as the gas, oil, and chemical industries, and above all, in nuclear power applications. As a result, more sophisticated statistical and probabilistic techniques for reliability life model development were developed on the computational side. These include maximum likelihood estimation (MLE) and Bayesian estimation techniques in reliability model developments. The explosive growth of the integrated circuits (IC) in the 1980s renewed interest in the PoF modeling. As a result, accelerated life testing became critical to reliability analysis as a means of understanding the underlying processes and failure mechanisms that cause the failure of ICs. Further, the use of environmental screening methods improved the IC and other complex device mass production. The 1980s also marked the development of initiatives for modeling dependencies at the system level. Most of these efforts tackled common cause failures as frequent dependency problems in systems. The common cause failure (CCF), which is the failure of more than one component due to a shared root cause, is classified as a dependent failure. Finally, during the 1980s, applications of reliability growth became helpful during the design of complicated systems, particularly in defense systems. Limited resources in conducting PoF resulted in limited test data. To account for the uncertainties in the PoF model, the 1990s experienced the development of the probabilistic physics of failure (PPoF) that would account for PoF model uncertainties, thus allowing for assignment measures of confidence over estimated reliability metrics (see Chatterjee and Modarres 2012). The 1990s also gave rise to the uses of time-varying accelerated tests (e.g., step-stress tests) and highly accelerated life testing (HALT) that introduced qualitative reliability information into the design process. The decade of the 2000s proved to be significant to reliability assessments that required powerful simulation tools. With the appearance of practical Monte Carlo simulation algorithms such as Markov chain Monte Carlo (MCMC) simulation, recursive Bayes, and particle filtering, condition-based monitoring (CBM) and prognosis and health management (PHM) became dominant topics in the field. Additionally,

4

Reliability and Risk Analysis

integrated logic and probabilistic models such as combinations of the traditional logic trees coupled with PoF, Bayesian networks or Bayesian belief networks (BN or BBN), and other methods from machine learning were introduced. Further extensions in big data, PHM, systems modeling, machine learning (ML), and related ML application fields had a commanding influence in the 2010s and beyond. Notably, complex regression models and formal accounting of uncertainties in reliability models, extensive causal Bayesian networks for modeling failure, semisupervised and unsupervised deep learning from failure data and sensors monitoring performance systems, application of reliability modeling to highly complex systems such as the cyber-physical systems, PoF-informed deep learning models of reliability have been examples advances in this decade. A significant thrust of research in PoF modeling was to rely on thermodynamics and information theory to explain failure mechanisms and move toward developing science-based PoF models instead of empirical models that have formed the foundation of these physical reliability prediction models. Bayesian networks have been used widely for modeling failure under changing conditions. A significant body of work on causal modeling has also emerged.

1.3 RELIABILITY MODELING APPROACHES The most widely accepted definition of reliability is the ability of an item (e.g., a product, component, or a system) to operate without failure under specified operating conditions to attain a mission having the desired length of time or number of cycles By dissecting this definition, we can see that to assess reliability correctly, we must do the following: 1. 2. 3. 4.

Define the item (i.e., part, component, subsystem, system, or structure) Define the mission and what constitutes success (or failure) Define the designated conditions and environments of use Specify a mission variable (e.g., time, cycles, stress level, and cumulative damage) 5. Assess ability (e.g., through testing, modeling, data collection, or analysis)

This process involves considerable engineering skills. Reliability engineering can be approached from a deterministic or probabilistic perspective. This book deals with the reliability analyses involving predicting failures using probabilistic notions. In a deterministic perspective, the primary view is that understanding failure mechanisms and the approaches to model them, such as physics of failure, and the resulting failure modes that may be associated with the design or that can be introduced from outside of the system (e.g., by users, maintainers, or the external operating environment) is sufficient. The probabilistic view builds on this by adding information about likelihood and uncertainty of events. Most failure mechanisms, interactions, and degradation processes in a particular design are generally hard to predict and sometimes not well understood. Accordingly, predicting specific failure modes or any functional failure involves uncertainty, and thus is inherently a probabilistic problem. Therefore, reliability prediction, whether

Reliability Engineering Perspective and Fundamentals

5

using physics of failure or relying on the historical occurrences of failures, is a probabilistic domain. It is the strength of the evidence and knowledge about the failure events and processes that describe failure mechanisms and associated failure modes and allow prediction of the reliability of the design and the corresponding confidence over that prediction. This book considers how to probabilistically assess the reliability and risk of multiple classes of items (units, systems, subsystems, structures, components, software, human operators, etc.). We will refer to two primary layers as components and systems. First, we discuss how things fail in the remainder of this chapter. Then, in Chapters 2–5, we will elaborate on how to conduct the reliability analysis of a component (i.e., an item for which enough information (data) is available to predict its reliability without deconstructing it further). Then, in Chapter 6, we will discuss methods to model the reliability of a system, which is a collection of multiple components whose proper, coordinated operation leads to the appropriate functioning of the system. Assessment of the reliability of a system based on its basic elements is one of the most important aspects of reliability analysis. In system reliability analysis, it is therefore imperative to model the relationship between various items (and their failures) and the reliability of the individual items themselves to determine the overall system’s reliability. The domain of coverage in reliability engineering is extensive. Engineering systems are composed of interacting hardware, software, humans, operational environments, and more (see Figure 1.1). The conditions and environment cannot be neglected, nor can the role of humans. When considering an engineering system, the same hardware design operating in different locations, under varied organizations, or cultures can have vastly different reliability outcomes. Imagine a vehicle operating in the desert vs. one operating in the arctic. Heat and dust cause different types of failures than cold and ice. Likewise, systems operated by experts will have lower failure rates than those operated by novices—one reason we enforce training requirements. Furthermore, organizational policies that enforce maintenance practices, monitor

FIGURE 1.1 Domains of coverage in reliability engineering.

6

Reliability and Risk Analysis

FIGURE 1.2

Framework for modeling failure.

failures, and invest in training the humans and software will inherently see high hardware reliability than those that neglect to monitor, maintain, and train. Because this book takes a probabilistic modeling approach to reliability, multiple facets of probability modeling are relevant. First, the data-driven statistical models are concerned with when and how often failures occur. Statistical models are used to predict static conditions where appropriate data exist. A second facet is causal modeling, concerned with why failures occur. These causal models enable us to analyze and predict changing (uncertain) conditions. Figure 1.2 depicts elements of a framework to construct failure models using capacity and challenge notions. Several simple failure models discussed by Dasgupta and Pecht (1991) follow the framework in Figure 1.2. A summary of these models has been provided below.

1.3.1

Physics of failure aPProach

Physics of failure views failures as occurring when challenges (e.g., operating stress, environmental stress, accumulated damage) exceed the capacity of the item to withstand the challenges. Both the challenges and the capacity may be affected by specific internal or external conditions. When challenges surpass the capacity of the system, a failure occurs. Or, viewing this in success space, reliability occurs when the system’s performance exceeds the specified requirement. Specific physics of failure models use different definitions and metrics for capacity and challenge. Adverse conditions generated artificially or naturally, internally, or externally, may increase or induce challenges to the system or/and reduce the item’s capacity to withstand the challenges. 1.3.1.1 Performance-Requirement Model In this model, an item’s performance (e.g., efficiency, flow output, reliability) is adequate if it falls within the acceptable or required limit (e.g., accepted margin, such as a safety margin or operational limit, design requirement, warranty life). Examples of this are a copier machine and a measuring instrument, two systems where gradual degradation eventually results in a user deciding that performance quality is unacceptable. Another example is the capability of an emergency safety injection pump at a nuclear power plant. When the pump’s output flow falls below a given required level, the pump may be considered functionally failed. The system’s performance may decline over time but must remain above the stated requirements (e.g., above the minimum flow output) to achieve its function (e.g., to safely cool a reactor if needed).

Reliability Engineering Perspective and Fundamentals

7

1.3.1.2 Stress-Strength Model In the stress-strength model, an item fails if the challenge (described by stress in the form of mechanical, thermal, etc.) exceeds the capacity (characterized by strength, such as yielding point, melting point, etc.). The stress represents an aggregate of the challenges and external conditions, often considered a sudden burst or overstressed, such as an item dropped on a hard surface. This failure model may depend on environmental conditions, applied loads, and the occurrence of critical events rather than the mere passage of time or cycles. Strength is treated either as a random variable (r.v.) showing stochastic variability or lack of knowledge about the item’s strength or as a known deterministic value. Stress may be represented as an r.v. Two examples of this model are (a) a steel bar in tension and (b) a transistor with a voltage applied across the emitter-collector. In this model, the item fails only when stress exceeds strength, and the memory of past stresses applied to the item is rarely considered. 1.3.1.3 Damage-Endurance Model This model is similar to the stress-strength model, but the scenario of interest is that stress (or load) causes damage over time that accumulates irreversibly, as in corrosion, wear, embrittlement, creep, and fatigue. The aggregate of these applied stresses leads to the metric represented as cumulative damage (e.g., total entropy, amount of wear, and crack length in case of fatigue-based fracture). The cumulative damage may not necessarily degrade performance; the item fails only when the cumulative damage exceeds the item’s endurance to damage (i.e., the damage accumulates until the endurance of the item is reached). As such, an item’s capacity is measured by its tolerance of damage or its endurance (such as fracture toughness). Unlike the stress-strength model, the accumulated damage does not disappear when the stresses causing the damage are removed; the stress leaves a memory of the past. However, sometimes treatments such as annealing reduce the amount of damage. In this modeling technique, damage and endurance may be treated as r.v.s. Like the stress-strength model, endurance is an aggregate measure for the effects of challenges and external conditions on the item’s ability to withstand cumulative stresses.

1.3.2

failure agents

In the physics of failure modeling approaches discussed in the previous three subsections, challenges are caused by failure-inducing agents (sometimes called failure agents). Examples of two of the most important failure-inducing agents are high stress and passage of time. High stress can be in mechanical (e.g., cyclic load), thermal (e.g., high heat, thermal cycling), electrical (e.g., voltage), chemical (e.g., salt), and radiation (e.g., neutron). For example, turning on and off a standby component may cause a thermal or mechanical load cycling to occur. The passage of time gives more opportunity for the normal environmental and operating stresses to cause small cumulative damage. Careful consideration of reliability requires analyzing the two failure-inducing agents (high stresses and simple passage time under normal stresses). To properly depict performance and requirements, it is necessary to understand why and how such conditions occur that reduce performance.

8

Reliability and Risk Analysis

The main body of this book addresses the probabilistic treatment of time or cycles to failure as agents of failure. Equally important, however, is understanding the failure mechanisms and their related models of physics of failure. It is also important to recognize that models representing failure mechanisms (related stress to damage or life) are also uncertain (i.e., their parameters are uncertain, and the model is associated with stochastic error due to failure data scatter). This means that the more appropriate model of failure mechanisms is the probabilistic physics of failure. For further readings on the physics of failure, see Modarres et al. (2017), Dasgupta and Pecht (1991), Pecht and Kang (2018), and Collins (1993).

1.4 DEFINITIONS Now that we understand the role that a well-established reliability analysis and engineering program can play in influencing the performance of items, let us define the important elements of the performance. The performance of an item is composed of four elements: • Capability: the item’s ability to attain the functional requirements. • Efficiency: the item’s ability to attain its functions economically and quickly with minimum waste. • Reliability: the item’s ability to start and continue to operate to fulfill a mission. • Availability: the item’s ability to become operational following a failure and remain operational during a mission. All these measures are influenced by the item’s design, construction, production, or manufacturing. Capability and efficiency reflect the levels to which the item is designed and built. For example, the designer ensures that the design meets the functional requirements. While reliability is influenced by design, reliability is also an operation-related performance metric influenced by the factors that enhance or degrade the ability to remain operational without failure. For a repairable item, the ease with which the item is maintained, repaired, and returned to operation is measured by its maintainability, which affects its availability. Capability, efficiency, reliability, and availability may be measured deterministically or probabilistically. For more discussions on this subject, see Modarres (2005). It is possible to have a highly reliable item that does not achieve high performance or capability. Examples include items that do not fully meet their stated design requirements. Humans play a significant role in the design, construction, production, operation, and maintenance of the item. These roles can significantly influence the values of the four performance measures. The role of the human is often determined by various programs and activities that support the four elements of performance, proper implementation of which leads to a quality item. Reliability and availability play a vital role in the overall performance. Therefore, of the four elements of performance discussed above, we are mainly interested in reliability and availability in this book. Reliability is a critical element of achieving high performance since it directly and significantly influences the item’s performance

Reliability Engineering Perspective and Fundamentals

9

FIGURE 1.3 Synergistic effects between risk and reliability of an item.

and, ultimately, its life cycle cost. Conversely, poor reliability causes increased warranty costs, liabilities, recalls, and repair costs. In this book, we are also interested in risk analysis. While the risk associated with an item is not a direct indicator of performance, a quantitative measure of risk can be an essential metric for identifying and highlighting items that contribute significantly to the risk and will be required to set and meet adequate performance levels for risk-significant items. Reliability and availability highly influence an item’s risk. For example, a highly reliable item is expected to fail less frequently, resulting in lower risk. On the other hand, the risk assessment of an item may be used to identify items that should attain high reliability or availability. Accordingly, the risk and reliability of an item synergistically influence each other. This concept is depicted in Figure 1.3.

1.4.1

reliability

The most widely accepted definition of reliability is the one in Section 1.3 repeated here for convenience: reliability is the ability of an item to operate without failure under specified operating conditions to attain a mission having the desired length of time or number of cycles. We typically model reliability as a function of use time or cycles of use. Now, let us expand this definition to include a mathematical description of reliability. Probabilistically, an item’s reliability can be expressed by R ( tinterest ) = Pr(Tfail > tinterest c1 ,  c2 , …),

(1.1)

where R ( tinterest ) is the reliability of the item at the time of interest tinterest , given c1, c2,... the designated operating conditions, such as the environmental conditions. Tfail denotes the r.v. of time to failure. The time of interest tinterest is the designated mission time (i.e., the length of time or number of cycles desired for the item’s operation). T could also represent the strength, endurance limit, or other quantities of interest. Clearly Tfail is most often an r.v., but rarely is a constant. tinterest is typically constant, but rarely an r.v.

10

Reliability and Risk Analysis

Often, in practice, c1, c2, … are implicitly considered in the probabilistic reliability analysis, and thus Equation 1.1 reduces to R ( tinterest ) = Pr ( Tfail > tinterest ).

1.4.2

(1.2)

risk

Risk is the item’s potential to cause a loss (e.g., loss to other systems, harm to individuals or populations, environmental damage, or economic loss). A related term is hazard, but this term is not interchangeable with risk. We define hazard as a source of damage, harm, or loss. Inherently, risk involves both uncertainty and loss/damage. Risk can be addressed both qualitatively and quantitatively. Qualitatively speaking, when there is a source of danger (e.g., a hazard) and imperfect safeguards to prevent the danger (e.g., prevent or mitigate exposure to the hazard), there is a possibility of loss or injury. Sometimes this possibility is quantified and represented probabilistically as a metric for risk. Risk can be more formally defined as the potential of loss (e.g., material, human, or environmental losses) resulting from exposure to a hazard. In complex engineering systems, there are often safeguards against exposure to hazards; safeguards may include physical barriers or preventive systems. The higher the level of safeguards, the lower the risk is. This also underscores the importance of highly reliable safeguard systems to prevent or mitigate losses and shows the roles and relationship between reliability analysis and risk analysis. An example of a risk metric could be the expected loss caused by an item during a mission or over a period. The expected loss as a measure of risk can also be viewed and assessed probabilistically. This book is concerned with quantitative risk assessment (QRA). Since quantitative risk assessment involves an estimation of the degree or probability of an uncertain loss, risk analysis is fundamentally intertwined with the concept of probability of occurrence of hazards. Formal risk assessment consists of answers to three questions (Kaplan and Garrick, 1981): 1. What can go wrong [that could lead to a hazard exposure]? 2. How likely is it to happen? 3. If it happens, what are the consequences are expected? A list of scenarios of events leading to the hazard exposures should be defined to answer question one. Then, the probability of these scenarios should be estimated (answer to question two), and the consequence of each scenario should be described (answer to question three). Therefore, the risk is defined, quantitatively, as the following triplet: R = Si , Pi , Ci ,  i = 1,2,…,  n,

(1.3)

where Si is a scenario of events that lead to hazard exposure, Pi is the probability or frequency of scenario i, and Ci is the consequence of scenario i, for example, a measure of the degree of damage or loss.

Reliability Engineering Perspective and Fundamentals

11

Since Equation 1.3 involves estimating the probability or frequency of occurrence of events (e.g., failure of protective, preventive, or mitigative safeguards), most of the methods described in Chapters 2–8 become relevant and useful. We have specifically devoted Chapter 9 to a more detailed, quantitative description of risk analysis. For more discussions on risk analysis, the reader is referred to Modarres (2006), Bedford and Cooke (2001), and Kelly and Smith (2011).

1.4.3

availability and Maintainability

Availability analysis is performed on repairable systems to ensure that an item has a satisfactory probability of being operational to achieve its intended objectives. An item’s availability can be thought of as a combination of its reliability and maintainability. Maintainability is an item’s ability to be quickly restored following a failure. Accordingly, when no maintenance or repair is performed (e.g., in nonrepairable items), reliability can be considered as instantaneous availability. Mathematically, the availability of an item is a measure of the fraction of time that the item is in operating condition as a function of either total system time or calendar time. There are several availability measures, namely, inherent availability, achieved availability, and operational availability. Here, we describe inherent availability, defined most commonly in the literature. A more formal definition of availability is the probability that an item, when used under stated conditions and support environment (i.e., perfect spare parts, personnel, diagnosis equipment, procedures, etc.), will be operational at a given time. Based on this definition, the mean availability of an item during an interval of time T can be expressed by A=

U ,  U+D

(1.4)

where U is the uptime during time T, D is the downtime during time T, and total time T = U + D. The mathematics and methods for reliability analysis discussed in this book are also equally applicable to availability analysis. Equation 1.4 can also be extended to represent a time-dependent availability. Time-dependent expressions of availability and availability measures for different types of equipment are discussed in more detail in Chapter 7. For further definition of these availability measures, see Ireson and Coombs (1988).

1.4.4

failure Modes and failure MechanisMs

Failure modes, failure mechanisms, and functions are building blocks of risk and reliability scenarios. The manner or way of the final failure occurrence is called the failure mode. Stated another way, a failure mode is a functional manner of item failure. An example of a failure mode for a valve would be external leakage. Failure modes answer the questions “How does the item fail?” or “What does the

12

Reliability and Risk Analysis

FIGURE 1.4 components.

Relationship between environmental conditions and failure mechanisms in

FIGURE 1.5

Relationship between item, functions, failure modes, and failure mechanisms.

item fail to do?” Typically, we assume that all item failure modes are mutually exclusive. From this discussion, we can conclude that the failure mode refers to a failed state of the item. Failure mechanisms are physical processes through which damage occurs. Damage can occur rapidly (abruptly) or slowly (cumulatively). Examples of failure mechanisms leading to rapid damage include brittle facture, melting and yielding. Examples of failure mechanisms causing cumulative damage are fatigue, wear and corrosion. In both cases failure occurs when the resulting damage exceeds the item’s capacity to withstand damage (e.g., strength, endurance, or required performance). An example of failure mechanism leading to external valve leakage (i.e., failure mode) would be corrosion. From this discussion, we can conclude that the failure mechanism refers to the process leading to the item’s failure mode (i.e., failed state). Figure 1.4 shows the relationship between environmental conditions and common failure mechanisms. The relationship between failure modes, failure mechanisms, and functions is illustrated in Figure 1.5.

1.5

SYSTEMS, FUNCTIONS, AND FAILURE

Assessment of the reliability of a system based on its basic elements is one of the most important aspects of reliability analysis. A system is a collection of items whose

Reliability Engineering Perspective and Fundamentals

13

coordinated operation achieves a specific function (or functions). These items may be subsystems, structures, components, software, algorithms, human operators, users, maintenance programs, and more. In reliability analysis, it is therefore imperative to model both the reliability of the individual items and the relationship between various items as they affect the system’s reliability. Several aspects must be considered in system reliability analysis. One is the reliability of the components that comprise the system. A component is a basic physical entity of a system analyzed from a reliability perspective (i.e., not further divided into more abstract entities). In Chapters 2–5, we elaborate on conducting the reliability analysis at a basic component level. A second consideration is the manner of the item’s failure (the item failure mode) and how it affects the system’s failure mode(s). Finally, another important consideration is the physical configuration of the system and its components. Beginning in Chapter 6, we discuss methods to model the relationship between components and systems, which allow us to determine overall system reliability. We will elaborate more on this in Chapter 6 when we explain methods and functions for modeling the reliability of a system with n components as a function of the reliability of those components. For example, consider a pumping system that draws water from two independent storage tanks—one primary and one backup tank, each equipped with a supply valve. If the supply valve on one tank fails closed, the pumping system can still operate because of the second tank. Thus, both the reliability of each component and the system configuration affects system reliability. Now, consider the effect of failure modes. If the tank’s supply valves fail in an open position, the pump can still draw water and successfully operate. However, if the supply valves fail in a closed position, the pump cannot work, and the pumping system fails. So, the failure mode of the valves is highly relevant for system reliability analysis. The line between component and system is arbitrary and varies depending on the objectives, scope, resources of modeling and analysis, and the state of the art and conventions. This relationship is conceptually shown in Figure 1.6. It is necessary to define and clearly articulate the elements and boundaries of the system in any system reliability analysis or data source. One example of a system definition is shown in Figure 1.7.

FIGURE 1.6

Notional relationship between components and systems.

14

Reliability and Risk Analysis

Fuel or Elec. Power

Driver (Diesel, El. Motor, etc,.)

Starting System

Control and Monitoring

Power

Exhaust

Inlet

Power Transmission (Gearbox, etc.)

Lubrication System

Remote

Pump Unit

Misc.

Coolant

Instr Boundary Pump System Power transmission • Gearbox/var. drive • Bearing • Seals • Lubrication • Coupling to driver • Coupling to driven unit • Instruments

Pump • • • • • • • • • • • •

Support Casing Impeller Shaft Radial bearing Thrust bearing Seals Valves & piping Cylinder liner Piston Diaphragm Instruments

Control and monitoring

Lubrication system

• Instruments • Cabling & junction boxes • Control unit • Actuating device • Monitoring • Internal power supply • Valves

• Instruments • Reservoir w/ heating system • Pump w/ motor • Filter • Cooler • Valves & piping • Oil • Seals

Miscellaneous • Purge air • Cooling/heat ing system • Filter, cyclone • Pulsation damper

FIGURE 1.7 Illustration of system boundaries (top) and system breakdown (bottom) for a pump. (Adapted from OREDA 2002.)

Reliability Engineering Perspective and Fundamentals

1.5.1

15

failure Modes and MechanisMs for Mechanical equiPMent

Some typical mechanical failure modes for active hardware components are: 1. 2. 3. 4. 5.

Premature operation, Failure to start operation when needed, Failure to continue operation after the start, Failure to stop operation at the specified time, Degraded operation.

A more detailed example list can be generated by expanding upon these failure modes. Many reliability data collection systems provide extensive taxonomies of failure modes. One example of failure modes for a pump can be found in Table 1.1. (Source: OREDA 2002): Now let’s turn to failure mechanisms. Mechanical failure mechanisms can be divided into damage-inducing, capacity-reducing, and a combination of both. Damage-inducing mechanisms refer to mechanisms that cause or result from localized damage, such as cracking. The damage may be temporary or permanent (cumulative). For example, elastic deformation may result from a force applied on the item that causes damage that disappears when the applied force is removed. However, cumulative damage caused by a failure mechanism is permanent. For example, fatigue is a mechanical failure mechanism whose direct effect is irreversible cumulative damage in the form of crack initiation and growth. TABLE 1.1 Detailed List of the Failure Modes of a Pump (OREDA, 2002) AIR BRD ERO ELP ELU FTS STP HIC INL LOO SER NOI OTH OHE PDE UST STD UNK VIB

Abnormal Instruments Reading Breakdown Erratic output External leakage—Process medium External leakage—Utility medium Fail to start on demand Fail to stop on demand High output Internal leakage Low output Minor in service problems Noise Other Overheating Parameter deviation Spurious stop Structural deficiency Unknown Vibration

16

Reliability and Risk Analysis

TABLE 1.2 Examples of Failure Mechanisms Damage-Inducing Wear Corrosion Cracking Diffusion Creep Fretting Fatigue

Capacity-Reducing Fatigue Embrittlement Thermal shock Diffusion Grain boundary migration Grain growth Precipitation hardening

Capacity-reducing mechanisms are those that lead (indirectly) to a reduction of the item’s strength or endurance to withstand applied stress or cumulative damage. For example, radiation may cause material embrittlement, thus reducing the material’s capacity to withstand cracks or other damage. Table 1.2 shows examples of failure mechanisms in each class of mechanisms. Table 1.3 summarizes the cause, effect, and physical processes involving common mechanical failure mechanisms.

1.5.2

failure Modes and MechanisMs for electrical equiPMent

Electrical failure mechanisms tend to be more complicated than those of purely mechanical failure mechanisms. This is caused by the complexity of the electrical items (e.g., devices) themselves. In integrated circuits, a typical electrical device, such as a resistor, capacitor, or transistor, is manufactured on a single crystalline chip of silicon, with multiple layers of various metals, oxides, nitrides, and organics on the surface, deposited in a controlled manner. Often a single electrical device comprises several million elements, compounding any reliability problem present at the single element level. Furthermore, once the electrical device is manufactured, it must be packaged with electrical connections to the outside world. These connections and the packaging are as vital to the proper operation of the device as the electrical elements themselves. Electrical device failure mechanisms are usually divided into three types: electrical stress failure, intrinsic and extrinsic failure mechanisms. These are discussed below. • Electrical stress failure occurs when an electrical device is subjected to voltage levels higher than design constraints, damaging the device and degrading electrical characteristics enough that the device may effectively fail. This failure mechanism is often a result of human error. Also known as electrical overstress (EOS), uncontrolled currents in the electrical device can cause resistive heating or localized melting at critical circuit points, which usually results in catastrophic failure but can also cause latent damage. Electrostatic discharge (ESD) is one common way of imparting large, undesirable currents into an electrical device.









Corrosion

Impact

Fatigue

Examples of Causes

Buckling

Mechanism

TABLE 1.3 Examples of Leading Mechanical Failure Mechanisms

1. Localized stresses 2. Deformation 3. Fracture





Examples of Effects

(Continued)

Failure or damage by the interaction of generated dynamic or abrupt loads that result in significant local stresses and strains Application of fluctuating loads (even far below the yield point) leading to a progressive failure phenomenon that initiates and propagates cracks

When load applied to items such as struts, columns, plates, or thin-walled cylinders reaches a critical value, a sudden significant change in geometry, such as bowing, winking, or bending, occurs Gradual damage of materials (primarily metals) by chemical and/or electrochemical reactions with the environment. Corrosion also interacts with other mechanisms such as cracking, wear, and fatigue

Description

Radiation damage



Thermal shock 1. Rapid cooling 2. Rapid heating 3. Sudden large differential temperature Yield











Creep

1. Progressive loss of material  2. Cumulative change in dimensions 3. Deformation

Examples of Effects





Examples of Causes

Embrittlement

Wear

Mechanism

TABLE 1.3 (Continued) Examples of Leading Mechanical Failure Mechanisms Wear is not a single process. It can be a complex combination of local shearing, plowing, welding, and tearing, causing the gradual removal of discrete particles from contacting surfaces in motion. Particles entrapped between mating surfaces. Corrosion often interacts with wear processes and changes the character of the surfaces Failure or damage from loss of material ductility (i.e., becoming brittle), often resulting from a combination of processing factors (e.g., cold or heat treatment), stress concentration points, and presence of hydrogen. Plastic deformation in an item accrues over some time under tensile, compressive, or bending stress until the accumulated dimensional changes interfere with the item’s ability to function properly. Thermal gradients in an item causing major differential thermal strains that, if exceeding the ability of the material to withstand, lead to failure and fracture. High enough tensile, compressive, or bending stress leads to plastic deformation in an item by operational loads or motion. Radiation causes rigidity and loss of ductility. Polymers are more susceptible than metals.

Description

19

Reliability Engineering Perspective and Fundamentals

• Intrinsic failure mechanisms are related to the electrical element itself. Most failure mechanisms related to the semiconductor chip and electrically active layers grown on its surface are in this category. Intrinsic failures are associated with the basic electrical activity of the device and usually result from poor manufacturing or design procedures. Thus, intrinsic failures cause both reliability and manufacturing yield problems. Common intrinsic failure mechanisms are gate oxide breakdown, ionic contamination, surface charge spreading, and hot electrons. In recent years semiconductor technology has reached a high level of maturity, with a correspondingly high level of control over intrinsic failure mechanisms. • Extrinsic failure mechanisms are external mechanisms for electrical devices that stem from device packaging and interconnections problems. For example, failures caused by an error during the design, layout, fabrication, or assembly process or by a defect in the fabrication or assembly materials are considered extrinsic failures. Also, a failure caused by a defect created during manufacturing is classified as extrinsic. Most extrinsic failure mechanisms are mechanical. However, embedded deficiencies or errors in the electronic device and packaging manufacturing process often cause these extrinsic mechanisms to occur, though the operating environment strongly affects the damage caused by these failure mechanisms. Because of recent reductions in damage accumulation through intrinsic failure mechanisms, extrinsic failures have become more critical to the reliability of the latest generation of electronic devices. Many electrical failure mechanisms are interrelated. A partial failure due to one mechanism can often evolve into another. For example, oxide breakdown may be caused by poor oxide processing during manufacturing, but it may also be exasperated by ESD, damaging an otherwise intact oxide layer. Likewise, corrosion and ionic contamination may be initiated when a packaging failure allows unwanted chemical species to contact the electronic devices. Then failure can occur through trapping, piping, or surface charge spreading. Also, intrinsic failure mechanisms may be initiated by an extrinsic problem once the package of an electrical device is damaged. There are a variety of intrinsic failure mechanisms that may manifest themselves in the device. Tables 1.4–1.6 summarize the cause, effect, and physical processes involving common electrical stress and intrinsic and extrinsic failure mechanisms. TABLE 1.4 Electrical Stress Failure Mechanisms Mechanism EOS

ESD

Causes Improper application of handling Common static charge buildup

Effect

Description

The device is subjected to voltages higher Localized melting Gate oxide breakdown than design constraints. Localized melting Contact with static charge buildup during Gate oxide breakdown device fabrication or later handling results in high voltage discharge into the device.

20

Reliability and Risk Analysis

TABLE 1.5 Intrinsic Failure Mechanisms Mechanism Gate oxide breakdown

Ionic contamination

Causes 1) EOS 2) ESD 3) Poor gate oxide processing 1) Introduction of undesired ionic species into the semiconductor

Effect

Surface charge spreading



Slow trapping



Hot electrons

1) High electric fields in the conduction channel

Piping

1.5.3

Description

1) Degradation in The oxide layer that separates current-voltage gate metal from the (I–V) characteristics semiconductor is damaged or degrades with time. 1) Degradation in I–V Undesired chemical species can characteristics be introduced to the device 2) Increase in through human contact, threshold voltage processing materials, improper packaging, and so on. Undesired formation of conductive pathways on surfaces alters the electrical characteristic of the device.

Defects at gate oxide interface trap electrons, producing undesired electric fields. High electric fields create electrons with sufficient energy to enter the oxide. Diffusion along crystal defects in the silicon during device fabrication causes electrical shorts.

huMan functions, failure, and reliability

Humans play many roles in engineering systems. Since the focus of modern reliability engineering is on the engineered system, not only the hardware, analyzing the reliability of humans (or, more appropriately, the human-machine teams (Groth et al., 2019)) that contribute to functioning of an engineered system is an essential part of reliability analysis. Human reliability analysis (HRA) is the aspect of reliability engineering that provides the methodologies to model these human failures. HRA is an embedded part of the reliability engineering process, not a separate process to be conducted independently. While in some aspects of reliability and risk analyses, human failures have been treated as a separate analysis, this unnecessarily fragments the reliability modeling of complex engineering systems. Reliability and risk analysis should treat human functions and failures as embedded in reliability analysis. The human-machine team is inherently another item considered as part of system reliability similar to the hardware. However, unlike many types of hardware components, humans can play multiple roles in a system—so they can be viewed as components achieving various functions at several levels of abstraction. Addressing human reliability within an engineered system requires techniques built upon those

21

Reliability Engineering Perspective and Fundamentals

TABLE 1.6 Extrinsic Failure Mechanisms Mechanism

Causes

Effect

Description

Packaging failures



1. Increased resistance 2. Open circuits

See section on mechanical failure mechanisms

Corrosion





Electromigration

1. High current densities 2. Poor device processing



Microcracks

1. Poorly processed oxide steps



Stress migration



1. Short circuits

Bonding failures





Die attachment failures





The combination of moisture, DC operating voltages, and ionic catalysts causes electrochemical movement of material, usually the metallization High electron velocities impact and move atoms, resulting in altered metallization geometry and, eventually, open circuits Poor interface control causes metallization to diffuse into a semiconductor. Often this occurs as metallic “spikes” The formation of a metallization path on top of a sharp oxide step results in a break in the metal or a weakened area prone to further damage Metal migration occurs to relieve high mechanical stress in the device Electrical contact to device package (bonds) are areas of high mechanical instability and can separate if the processing is not strictly controlled Corrosion or poor fabrication causes voids in die-attach or partial or complete deadhesion

Particulate contamination

1. Poor manufacturing and chip breakage

1. Short circuits

Radiation



1. Various degrading effects

Contact migration



Conductive particles may be sealed in a hermetic package or may be generated through chip breakage High energy radiation (ionizing or cosmic) can create hot electronhole pairs that can interfere with and degrade device performance

22

Reliability and Risk Analysis

used in component and systems analysis. Thus, HRA is one of the special topics covered in Chapter 8.

1.6 PUTTING IT ALL TOGETHER: RISK ASSESSMENT MODELING As we conclude this chapter, let’s briefly discuss the general steps in risk assessment, which provides an integrated framework for explaining how the book fits together (Figure 1.8) as it builds toward Chapter 9. The key elements of a risk assessment include: 1. Define system and assemble information. a. Define the system to be analyzed and its components, interfaces, and boundaries, b. Collect, analyze, and model component failure data. 2. Define the problem scope objectives and select the appropriate methodology. a. Identify hazards, operating modes, and/or initiating events of interest. 3. Construct causal models and conduct causal analysis. a. Scenario development and modeling, b. Identify root causes (failure modes and failure mechanisms if needed), c. Identify failure effects, d. Construct logic models. 4. Qualitatively evaluate those models. a. System failure logic. 5. Integrate logic models, a. For example, many nuclear probabilistic risk assessments (PRAs) use a combination of event trees and fault trees. For complex systems, it is common to build separate models for different functions or subsystems or separate parts of the analysis (e.g., hardware vs. human reliability). 6. Evaluate consequences, a. Use a variety of models to assess effects and outcomes of the scenarios.

FIGURE 1.8 A simplified view of the process of risk assessment forms the framework of this book.

Reliability Engineering Perspective and Fundamentals

23



For Step (1), we discussed how to define (and document) the system to be analyzed in Chapter 1, (or we can either add content about or find references to techniques such as functional breakdown and structural breakdown, structural and functional block diagrams, and more). Next, data collection and analysis will be discussed in Chapters 4 and 5. For Step (2), several objectives are possible, and knowing your objectives will help you select the appropriate methodology. Objectives can range from design insight, demonstration of risk acceptability or tolerability, regulatory or code compliance, operational design support, system health management, and more. In Chapter 1, we define critical terms used establishing objectives and methods. The remainder of this textbook discusses a range of methodologies to support this step. For Step (3), in this chapter, we discuss failure mode and mechanism taxonomies. Chapter 6 will expand upon this and connect to identifying failure causes and effects through our discussion of several qualitative and quantitative models. Chapter 6 also discusses the construction of the qualitative logic models. For Steps (4) and (5), we will discuss the integration and evaluation of the logic models within Chapter 6 and Chapter 9. See Moradi and Groth (2020) for discussion of integrating additional types of information. Step (6) typically falls into the domain of deterministic modeling or loss calculations. It may include use of multiphysics models or detailed fire behavior models. While full coverage of this falls outside the scope of this book, Chapter 9 illustrates the use of fire modeling within this process via the HyRAM methodology (Groth and Hecht, 2017). For Step (7), we discuss Step 7a extensively in Chapters 3 and 5 (for nonrepairable items), Chapter 7 (for repairable items) and Chapter 8 (for common cause failure analysis, importance ranking, uncertainty and sensitivity analysis and the human reliability). Step 7b is explained in Chapter 6. The remaining aspects of Step (7) are advanced topics beyond the scope of this first course in reliability engineering—each of these nuanced topics is covered in more depth in advanced books such as Modarres (2006). For Step (8), see Modarres (2006). Finally, Step (9) is an important reminder that this is an iterative process, and not a one-time activity that we place on the shelf after the system is built. Some aspects of this step are presented in Chapter 9. This step is described in more detail in Modarres (2006).

24

Reliability and Risk Analysis

This book culminates in Chapter 9, which presents the integration of this process and presents several examples.

1.7 EXERCISES 1.1 Explain the difference between a failure mechanism and a failure mode. 1.2 Discuss the relationship between reliability and failure, both in words and mathematically. 1.3 Select a failure mechanism and find a news article where it has been relevant to an engineering system. 1.4 Select a failure mechanism and explain what activates or causes it to happen, what is the process through which damage occurs, what is the nature of the damage, what is the capacity (strength or endurance) to the damage, and what failure mode(s) can result at the end. 1.5 Describe the relationship between reliability and availability. 1.6 Select an accident or failure that you are familiar with. Write a failure scenario narrative containing the following: a. A brief description of the system and the accident or failure and what initiates the accident. b. The system failure mode, operating environment, and health, safety or environmental consequences. c. At least one of each of the following: a hardware failure mode, a failure mechanism and influencing factor contributed to this system failure. d. A human activity that contributed to the failure. e. A relevant image. 1.7 Read about the accident at a chemical plant in Bhopal, India. Describe what are the three fundamental elements of the risk triplet. 1.8 An item is down an average of 16 hours a year for repair and maintenance. What is the item’s mean availability? 1.9 Described five critical information that reliability analysis provides as input to a risk assessment. 1.10 Graphically and conceptually show the deterministic stress/damage vs. capacity (e.g., strength and endurance) for the three PoF models. Repeat the graphs to conceptually depict the three PoF models probabilistically.

REFERENCES Bedford, T. and R. Cooke, Probabilistic Risk Analysis: Foundations and Methods. Cambridge University Press, London, 2001. Chatterjee, K. and M. Modarres. “A Probabilistic Physics-of-Failure Approach to Prediction of Steam Generator Tube Rupture Frequency.” Nuclear Science and Engineering, 170(2), 136–150, 2012. Collins, J. A., Failure of Materials in Mechanical Design, Analysis, Prediction, and Prevention. 2nd edition, Wiley, New York, 1993. Dasgupta, A. and M. Pecht, “Materials Failure Mechanisms and Damage Models.” IEEE Transactions on Reliability, 40 (5), 531–536, 1991.

Reliability Engineering Perspective and Fundamentals

25

Groth, K. M. and E. S. Hecht, HyRAM: “A Methodology and Toolkit for Quantitative Risk Assessment of Hydrogen Systems.” International Journal of Hydrogen Energy, 42, 7485–749, 2017. Groth, K. M., R. Smith, and R. Moradi. “A Hybrid Algorithm for Developing Third Generation HRA Methods Using Simulator Data, Causal Models, and Cognitive Science.” RWeliability Engineering & System Safety, 191, 106507, 2019. Ireson, W. G. and C. F. Coombs, eds., Handbook of Reliability Engineering and Management. McGraw-Hill, New York, 1988. Kaplan, S. and J. Garrick, “On the Quantitative Definition of Risk.” Risk Analysis, 1(1), 11–27, 1981. Kelly, D. L. and C. Smith, Bayesian Inference for Probabilistic Risk Assessment: A Practitioner’s Guidebook. Springer -Verlag, London, 2011. Modarres, M., “Technology-Neutral Nuclear Power Plant Regulation: Implications of a Safety Goals-Driven Performance-Based Regulation.” Nuclear Engineering and Technology, 37(3), 221–230, 2005. Modarres, M., M. Amiri, and C. R. Jackson, Probabilistic Physics of Failure Approach to Reliability: Modeling, Accelerated Testing, Prognosis and Reliability Assessment. John Wiley, New York, 2017. Modarres, M., Risk Analysis in Engineering. CRC Press, Boca Raton, FL, 2006. Moradi, R. and K. M. Groth, “Modernizing Risk Assessment: A Systematic Integration of PRA and PHM Techniques.” Reliability Engineering & System Safety, 204, 107194, 2020. OREDA. Offshore Reliability Data Handbook, 4th edition, OREDA Participants, Høvik, Norway 2002. Pecht, M. and M. Kang, eds., Prognostics and Health Management of Electronics: Fundamentals, Machine Learning, and the Internet of Things. John Wiley & Sons, New York, 2018.

2

Basic Reliability Mathematics Probability

2.1

INTRODUCTION

In this chapter, we discuss the elements of probability theory relevant to the study of reliability of engineered systems. Probability is a concept that people use formally and casually every day. Weather forecasts are probabilistic in nature. People use probability in their casual conversations to show their perception of the likely occurrence or nonoccurrence of particular events. Odds are given for the outcomes of sport events and are used in gambling. The formal use of probability concepts is widespread in physics, astronomy, biology, finance, and engineering. We begin with a presentation of basic concepts of probability and Boolean algebra used in reliability analysis. The emphasis in this chapter is on presenting these fundamental mathematical concepts which are used and expanded upon in the remaining chapters in this book. For some readers, this chapter is a review of mathematics that they have learned in previous courses. Some details may be omitted. It is worth reviewing this chapter briefly to see the established terminology, notation, and key assumptions used throughout the book.

2.2

EVENTS AND RANDOM VARIABLES USED IN RELIABILITY

There are two basic units of analysis involved in the probabilistic approach to reliability: events and random variables. An event, E, is an outcome or set of outcomes of a process. An event occurs with some probability. The sample space, S, is a mutually exclusive and collectively exhaustive list of all possible outcomes of a process. For example, examining the outcomes of rolling a die, the sample space contains events S = {1,2,3,4,5,6}. Thus, an event is a combination of one or more sample points of interest. For example, the event of “an odd outcome when rolling a die” represents the event space containing outcome points 1, 3, and 5. In reliability engineering, E could be event of a pump failure, or that a pump fails to start on demand, or that multiple pumps fail simultaneously. One use of probability is to communicate the chance of various combinations of events, such as multiple components failing simultaneously in a specific period of time. We may also be concerned with characterizing random variables (r.v.) associated with events. For instance, for an event of component failure, the r.v. X could be an r.v. representing time to failure of a component, the number of failures that occur by a given time, the number of cycles until a first failure, the number of failure DOI: 10.1201/9781003307495-2

27

28

Reliability and Risk Analysis

components drawn from a population, the failure rate, the rate of occurrence of failures, and more. The lowercase x could be the probability of the event of interest. Random variables can be divided into two classes, namely discrete and continuous. An r.v. is said to be discrete if the values it can take are finite and countable, such as the number of system breakdowns in 1 year. An r.v. is said to be continuous if it can take on a continuum of values. It takes on a range of values from an interval(s) as opposed to a specific countable number. Continuous r.v.s result from measured variables as opposed to counted data. For example, the failure time of a light bulb can be modeled by a continuous r.v. time, T, which can take on a value within a continuous range. For any specific light bulb, we can observe its failure time ti. In this book, we will use an uppercase letter (e.g., X, Y) to represent an event or an r.v. We will use a lowercase letter to denote the value that the r.v. can take. For example, if X represents the number of system breakdowns in a process plant, then xi is the actual number of observed breakdowns during a given time interval.

2.3

SETS AND BOOLEAN ALGEBRA

In reliability engineering, we are analyzing specific events, data, or combinations. A set is a collection of items or elements, each with some specific characteristics. The elements of the set are contained in brackets {}. We use sets to talk about combinations or items of interest, e.g., data or events. For example, we may have a population of light bulb failure times representing early failures (e.g., those light bulbs with failure times less than 100 hours). Or a set containing two pumps, A and B. Suppose event A is the event that pump A fails to start, and event B is the event that pump B fails to start. For the event that both pumps fail to start—this event would be the set consisting of A and B. A set that includes all possible items of interest is called a universal set, denoted by Ω (which sometimes is replaced by 1 in engineering notation). A null set or empty set, φ (or 0 in some engineering notations), refers to a set that contains no items. A subset, denoted by ⊂ or ⊆, refers to a collection of items that belong to another set. For example, if set Ω represents the collection of all pumps in a power plant, then the collection of auxiliary feedwater pumps, A, is a subset of Ω, denoted as A ⊂   Ω. Graphically, the relationship between subsets and sets can be illustrated through Venn diagrams. The Venn diagram in Figure 2.1 shows the universal set Ω by a rectangle and subsets A, B, and C by circles. It can also be seen that B is a subset of A. The relationship between subsets A and B and the universal set can be symbolized by B ⊂ A ⊂ Ω. In this diagram, A is a superset of B, denoted as A ⊃ B. Subsets A and C Ω

A B

C

FIGURE 2.1 Venn diagram showing the relationship between subsets A, B, C and universal set Ω.

29

Basic Reliability Mathematics

have elements in common, but B and C do not. In this case B and C are called disjoint or mutually exclusive events. The complement of a set A, denoted by A and called A not, represents negation. It is the set of all events in the universal set that do not belong to subset A. In Figure 2.1, the non-shaded area outside of the set B bounded by the rectangle represents B. The sets B and B together comprise Ω. One can easily see that the complement of a universal set is a null set, and vice versa. That is, Ω = ∅ and Ω = ∅. The union of two sets, A with B, is a set that contains all items that belong to A or B or both. The union is symbolized either by A ∪ B or by A + B and is read “A or B.” The shaded area in Figure 2.2 shows the union of sets A or B. Suppose A and B represent odd and even numbers between 1 and 10, respectively. Then A = {1,3,5,7,9} , and B = {2, 4, 6, 8, 10} . Therefore, the union of these two sets is A ∪ B = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10}. If A = {x, y, z} and B = {t, x, z}, then A ∪ B = {t ,  x ,  y,  z } . Note that elements x and z are in both sets A and B but appear only once in their union. The intersection of two sets, A and B, is the set of items that are contained in both A and B. The intersection is symbolized by A ∩ B or A· B, which is usually simplified to AB. The intersection is read “A and B.” In Figure 2.3, the shaded area represents the intersection of A and B. If we have discrete sets A = {x, y, z} and B = {t, x, z}, then A ∩ B = { x ,  z}. Now, consider a set representing continuous items. Suppose A is a set of manufactured devices that operate for t > 0 but failed before 1,000 hours of operation. If set B represents a set of the devices that operated between 500 and 2,000 hours, then A ∩ B represents devices that have worked between 500 and 1,000 hours and belong to both sets and can be expressed as follows: A = {t   0 < t < 1,000},  B = {t   500 < t < 2,000}, A ∩ B = {t   500 < t < 1,000}. In two mutually exclusive or disjoint sets, A and B, A ∩ B = ∅. In this case, A and B have no elements in common. Two mutually exclusive sets are illustrated in Figure 2.4. Ω A

FIGURE 2.2

B

Union of two sets, denoted A or B, ( A ∪ B ). Ω A

FIGURE 2.3

B

Intersection of two sets, denoted A and B, ( A ∩ B ).

30

Reliability and Risk Analysis Ω A

B

FIGURE 2.4 Mutually exclusive sets A and B, ( A ∩ B ) = ∅.

From the discussions thus far, and from the examination of the Venn diagram, these conclusions can be drawn: The intersection of set A and a null set is a null set: A ∩ ∅ = ∅.

(2.1)

The union of set A and a null set is A: A ∪ ∅ = A.

(2.2)

The intersection of set A and its complement A is a null set: A ∩ A = ∅.

(2.3)

The intersection of set A and a universal set is A: A ∩ Ω = A.

(2.4)

The union of set A and a universal set is the universal set: A ∪ Ω = Ω.

(2.5)

The complement of the complement of set A is A: A = A.

(2.6)

A ∪ A = A.

(2.7)

The union of two identical sets A is A:

The intersection of two identical sets A is A: A ∩ A = A.

(2.8)

31

Basic Reliability Mathematics

TABLE 2.1 Boolean Algebra Laws Designation

Mathematical Notation

Engineering Notationa

Identity laws

A∪∅ = A A∪Ω = Ω A∩∅ = ∅ A∩Ω = A

A+0 = A A +1= 1 A⋅0 = 0 A ⋅1 = A

Idempotent laws

A∩ A = A A∪ A = A

A⋅ A = A A+ A= A

Complement laws

A∩ A = ∅ A∪ A = Ω

A⋅ A = 0 A⋅ A = 1

Law of absorption

A ∪ ( A ∩ B) = A A ∩ ( A ∪ B) = A

A + ( A ⋅ B) = A A ⋅ ( A + B) = A

de Morgan’s theorem

( A ∩ B) = A ∪ B ( A ∪ B) = A ∩ B

( A ⋅ B) = A + B ( A + B) = A ⋅ B

Commutative laws Associative laws

A∩ B = B∩ A A∪ B = B∪ A

A⋅ B = B⋅ A A+ B = B+ A

A ∪ ( B ∪ C ) = ( A ∪ B) ∪ C = A ∪ B ∪ C A ∩ ( B ∩ C ) = ( A ∩ B) ∩ C = A ∩ B ∩ C

A + ( B + C ) = ( A + B) + C = A + B + C A ⋅ ( B ⋅ C ) = ( A ⋅ B) ⋅ C = A ⋅ B ⋅ C

Distributive laws

A ∩ ( B ∪ C ) = ( A ∩ B) ∪ ( A ∩ C ) A ∪ ( B ∩ C ) = ( A ∪ B) ∩ ( A ∪ C )

A ⋅ ( B + C ) = ( A ⋅ B) + ( A ⋅ C ) A + ( B ⋅ C ) = ( A + B) ⋅ ( A + C )

a

In this table, we use A ⋅ B to denote Boolean intersection. In the remainder of the book, we simplify this to AB.

These identity laws and idempotent laws are a few axioms of Boolean algebra. Boolean algebra provides a means of evaluating sets. The rules are fairly simple. The axioms in Table 2.1 provide the major relations of interest in Boolean algebra. Example 2.1 Simplify the expression: E ∩ ( E ∩ Ω ).

Solution: Since E ∩ Ω = E and E ∩ E = E, then the expression reduces to E.

32

Reliability and Risk Analysis

Example 2.2 Simplify the following Boolean expression:

) (

(

)

( A ∩ B ) ∪ A ∩ B ∪ A ∩ B   

(

) (

)

de Morgan's theorem

) (

) (

)

Complementation law

= ( A ∩ B) ∩ A ∩ B ∩ A ∩ B

(

de Morgan's theorem

= A∪ B ∩ A∪ B ∩ A∪ B

{ ( ) ( )} = {  A ∪ ( B ∩ A ) ]∪[ ( B ∩ A ) ∪ ( B ∩ B ) } ∩ ( A ∪ B ) =  A ∩ A ∪ B ]∪[ B ∩ A ∪ B  ∩ ( A ∪ B )

(

)

Distributive law Distributive law (twice)

=  A ∪ B ∩ A  ∩ ( A ∪ B )

Absorption law

= A ∩ ( A ∪ B)

Absorption law

= A∩ B Note that other solutions are possible. The reader is encouraged to explore additional solutions to this example.

2.4 PROBABILITY TERMINOLOGY AND INTERPRETATIONS In using probability theory, we are concerned with assigning a probability to an event. So, what is probability? While this question seems simple, it is one with multiple answers, owing to different schools of thought regarding how to interpret probabilities and what data underpin them. See Cox (1946) for further discussion. Generally, we define probability as a numerical measure of the chance that an event occurs or more generally that a hypothesis is true. It is used to quantitatively express the uncertainty about the occurrence of (or outcome of) an event. A closely related term is frequency, which is defined as the rate of occurrence of events, i.e., the number of times an event occurs over a given period of time (or space or number of trials). To differentiate the concepts of probability and frequency, consider a cloud server with an average frequency of failure of 5 × 10 −2 failures per year. This means that the average time between failures will be 20 years, not that a failure will occur exactly at 20 years operation. Frequency can also be used to determine the probability of an event per unit of time, and/or to compute the frequency of multiple, jointly occurring events (e.g., multiple system failures). Likelihood as a quantitative measure should be differently interpreted than probability. Generally, whereas probability is about an event, or more generally, a proposition given available evidence or observed data, likelihood describes the probability of observing the evidence or data given an event or a proposition. For example, consider if one flips a coin three times and gets the data {Head, Tail, Head}. Given this observation, you may ask, “what is the likelihood that the proposition that the coin

33

Basic Reliability Mathematics

is fair is true?” Reversely, by assuming the truth of the proposition that the coin is fair, you may ask, “what is the probability that in three flips of a fair coin we get {Head, Tail, Head}?” Note that these questions yield the same numerical answer, but the interpretation of that number is quite different between the likelihood and probability. We will further elaborate on the more specific interpretations of probability and likelihood from two schools of thought—namely frequentist and Bayesian interpretations. While probability theory is concerned with determining the probability of events, statistics is the field of study concerned with collecting, organizing, analyzing, presenting, and interpreting data. Statistics applies probability theory to draw conclusions from data. For example, to draw conclusions about the likelihood that the coin is fair in the example above, we use statistics to evaluate the data and reach the conclusion. The probability will be associated with an event of interest or a discrete random variable E. We denote this as Pr ( E = e ) or simply Pr ( e ), both of which indicate the probability that random variable E takes on value e. We may also refer to this generally as Pr ( E ) denoting some event of interest which may take on states E1  , …, En. It is important that the reader appreciates the intuitive differences between the three major conceptual interpretations of probability, all of which are used in different aspects of reliability engineering.

2.4.1

classical interPretation of Probability (equally likely concePt, or saMPle sPace Partitioning)

In this interpretation the probability of an event, E, can be obtained from partitioning the sample space, S, provided that the same space contains n equally likely and different outcomes. This interpretation gives rise to the classical interpretation of probability. However, this is rarely the case in engineering. The probability of any event E describes a comparison of the relative size of the subset represented by the event E to the sample space S. It can be obtained from the following equation: Pr ( E ) =

ne , n

(2.9)

where n denotes the number of elements in the sample space S, of which ne have an outcome (event) E. For example, the probability of the event “rolling an odd number on a die” is determined by using: sample space n = 6, and ne ( odd   outcomes ) = 3. 3 Here, Pr ( odd  outcomes ) = = 0.5. 6 This definition is often inadequate for engineering applications. For example, if failures of a pump in a process plant are observed, it is unknown whether all failure modes are equally likely to occur. Nor is it clear if the whole spectrum of possible events is observed. When all sample points are not equally likely to be the outcome, the sample points may be weighted according to their relative frequency of occurrence over many trials or according to subjective judgment.

34

2.4.2

Reliability and Risk Analysis

frequency interPretation of Probability

In this interpretation, the limitation of the lack of knowledge about the overall sample space is remedied by defining the probability as the relative number of occurrences of an event in a large number of identical and independent trials—that is the limit of ne/n as n becomes large. Therefore, n Pr ( E = e ) = lim  e  . n  →   ∞    n 

(2.10)

This interpretation provides an empirical method to estimate probabilities by counting over the ensemble of many trials. Thus, if we have observed n = 2,000 starts of a pump in which ne = 20 failed, and if we assume that 2,000 is a large number and these pumps are identical, independent and exchangeable, then the probability of the 20 pump’s failure to start is = 0.01. 2,000 In the frequentist interpretation the idea of a large ensemble of trials of identical, independently distributed, and exchangeable events is vital. In some cases, there is a natural ensemble such as tossing a coin or repeated failures of the same equipment over many identical and independent trials reflecting identical use cases. However, this interpretation does not cover cases in which little or no data is available, nor cases where estimates are intuitive. Thus, a broader definition of probability is required, which has led to the subjective interpretation of probability.

2.4.3

subjective interPretation of Probability (bayesian Probability; evidential Probability)

In many problems involving one-of-a-kind events (e.g., the weather at a specific location on a specific date), or when limited data and information are available (e.g., for new technologies), definition and creation of a large set of independent, identical, and interchangeable events is impossible. For example, consider the probability of the existence of intelligent life on another planet. To implement the frequentist interpretation requires creating an ensemble of identical, interchangeable planets that would be purely hypothetical. Another example might be the performance of a particular student on an upcoming exam in this class; while we can assess evidence-based probabilities of past performance of a population of students on the exam, any particular student would balk at considering their exam grade as identical and interchangeable with another student. When there is no historical data, a small population, or other limitations the frequency interpretation becomes inappropriate and we instead use the subjective, Bayesian, or evidential interpretation of probability. In this interpretation, which can be traced back to Bernoulli and Laplace, a probability can be assigned to any statement as a way of representing possibility or the degree to which the statement is supported by the available evidence. That is, the probability Pr(E) is interpreted as the degree of belief (or confidence) of an individual in the truth of a proposition or event E. It is a representation of an individual’s state of knowledge, and when using it we generally assume that any two rational individuals with the same knowledge, information, and beliefs will assign the same probability.

35

Basic Reliability Mathematics

These probabilities are referred to as Bayesian because of the central role they play in the uses of Bayes’ Theorem. Different individuals may have dissimilar degrees of belief about the truth of a proposition, but they base those beliefs and can change their individual beliefs, based on introspection or revelation and new data and information. To better understand this interpretation, consider the probability of improving a system by making a design change. The designer believes that such a change will result in a performance improvement in one out of three missions in which the system is used. It would be difficult to describe this problem through the classical or frequentist interpretations. The classical interpretation is inadequate since there is no reason to believe that performance is equally likely to improve or not improve. The frequency interpretation is not applicable because no historical data exist to show how often a design change resulted in improving the system. Thus, the subjective interpretation provides a broad definition of the probability concept. Clearly, the subjective interpretation provides a consistent and rational method for reasoning about unique and singular events; it is not limited to repeatable and identical events (but can also be applied to those events). While the subjective component of interpretation cannot be entirely eliminated, the calibration reduces this subjectivity. The Bayesian rules for processing information and updating probabilities (as discussed in the next section) are themselves objective. Accordingly, new data and information can be objectively combined into posterior probabilities that allow one to continuously suppress the subjectivity in the probabilities elements as we consider increasingly more information. Note that the way a frequentist view looks at the data and proposition differs from a subjectivist view. Whereas a frequentist view treats the proposition as fixed and the data as random, the subjectivist view treats the truth of a proposition as random and places a degree of belief (or certainty) on its truth considering the fixed observed data or evidence. We will further discuss this in Chapter 4.

2.5

LAWS AND MATHEMATICS OF PROBABILITY

The basic mathematics underpinning probability are the same, regardless of the interpretation of probability.

2.5.1

definitions

The marginal (or unconditional) probability, Pr(A) is the probability of event A occurring. The joint probability of two events, A and B, is the probability that both events occur and is denoted as Pr ( A ∩ B ) , Pr ( A,  B ) , or  Pr ( A ⋅ B ). The conditional probability Pr ( A B ) is the probability of event A occurring given that B has occurred. This is read as “the probability of A, given that B has occurred.” The conditional probability is a function of the joint and marginal probabilities: Pr ( A B ) =

Pr ( A ∩ B ) . Pr ( B )

(2.11)

36

Reliability and Risk Analysis

Note that conditional probability is not commutative, i.e., Pr ( A B ) ≠ Pr ( B A ) except under the special condition that Pr(A) = Pr(B). It is important to understand two additional terms that are sometimes confused: mutually exclusive and independent. Recall the concept of mutually exclusive events from Section 2.3. Two events are mutually exclusive when A ∩ B = ∅, i.e., they have no elements in common. This means that A and B cannot happen simultaneously. For example, if A = rolling an even number on a die, and B = rolling an odd number on a die. Another example of mutually exclusive events is: event E1 = engine is in a failed state and event E2 = engine is in an operational state. By definition, the probability of occurrence of two mutually exclusive states is zero: Pr ( A ∩ B ) = 0.

(2.12)

By contrast, two events are independent, which is denoted as A ⊥ B, if the occurrence or nonoccurrence of one does not depend on or change the probability of the occurrence of the other. For example, if event A is rolling the number one on the first roll of a die, and event B is rolling a one on the second roll of a die. Another example could be event A being a valve failing to open and event B being an unrelated engine failing to start. This is expressed as Pr ( A B ) = Pr ( A ).

(2.13)

It is important to emphasize the difference between independent events and mutually exclusive events, since these two concepts are sometimes confused. In fact, two events that are mutually exclusive are not independent. Since two mutually exclusive events A and B have no intersection, that is, A ∩ B = ∅, then Pr ( A ∩ B ) = Pr ( A )·Pr ( B A) = 0. This means that Pr( B A) = 0, since Pr ( A ) ≠ 0. For two independent events, we expect to have Pr( B A) = Pr ( B ), which is not zero except for the trivial case of Pr ( B ) = 0. This indicates that two mutually exclusive events are indeed dependent. An example to illustrate this would be flipping a coin. In this scenario there are only two possible outcomes, a head or a tail, but not both, hence the events are mutually exclusive. However, if a head is obtained, then the probability of the tail is zero. Hence, these two events are not independent. Example 2.3 Let’s consider the result of a test on 200 manufactured identical parts. Let L represent the event that a part does not meet the specified length and H represent the event that the part does not meet the specified height. It is observed that 23 parts fail to meet the length limitation imposed by the designer and 18 fail to meet the height limitation. Additionally, seven parts fail to meet both length and height limitations. Therefore, 152 parts meet both specified requirements. Are events L and H dependent?

Solution: According to Equation 2.9, Pr(L) = (23 + 7)/200 = 0.15 and Pr(H) = (18 + 7)/200 = 0.125. Furthermore, among 25 parts (18 + 7) that have at least event H, seven parts

37

Basic Reliability Mathematics also have event L. Thus, Pr(L|H) = 7/25 = 0.28. Since Pr(L|H) ≠ Pr(L), events L and H are dependent.

2.5.2

axioMs of Probability and their iMPlications

The axioms of probability, or Kolmogorov axioms, are defined as: 1. Pr ( Ei ) ≥ 0, for every event Ei .

(2.14)

2. Pr ( Ω ) = 1.

(2.15)

3. Pr ( E1 ∪ E2 ∪…∪ En ) = Pr ( E1 ) + Pr ( E2 ) + …+ Pr ( En ) , when the events E1 , …,  En are mutually exclusive.

(2.16)

By examination of these axioms, there are several implications that become evident. First, that probabilities are numbers that range from 0 to 1: 0 ≤ Pr ( Ei ) ≤ 1.

(2.17)

The probability of the null set is 0, Pr ( ∅ ) = 0.

(2.18)

Pr ( A ) ≤ Pr ( B ) .

(2.19)

We can also state that if A ⊂ B then

Finally, the probability of a complement of an event is 1 minus the probability of that event:

( )

Pr Ei = 1 − Pr ( Ei ).

2.5.3

(2.20)

MatheMatics of Probability

A few rules are necessary to enable manipulation of probabilities. By rearranging the definition of conditional probability (Equation 2.11), we can see that the joint probability of two events can be obtained from the following expression: Pr ( A ∩ B ) = Pr ( B A ) Pr ( A ).

(2.21)

The chain rule of probability or the multiplication rule is the generalized form of this relationship. It gives the joint occurrence of n events E1 , …, En as a product of conditionals: Pr( E1 ∩ E2 ∩…∩ En ) = Pr ( E1 ) ⋅ Pr( E2 |E1 ) ⋅ Pr ( E3 |E1 ∩ E2 )… Pr ( En |E1 ∩ E2 ∩ …∩ En −1 )

(2.22)

38

Reliability and Risk Analysis

where, Pr( E3 | E1 ∩ E2 ∩ …) denotes the conditional probability of E3, given the occurrence of both E1 and E2, and so on. It is easy to see that when A and B are independent and Equation 2.13 is applied, Equation 2.21 reduces to: Pr ( A ∩ B ) = Pr ( A ) Pr ( B ) .

(2.23)

And thus, if all events are independent (i.e., E1 ⊥ E2 ⊥ … ⊥ … En) the chain rule of probability simplifies to: Pr ( E1 ∩ E2 ∩…∩ En ) = Pr ( E1 ) ⋅ Pr ( E2 )…⋅ Pr ( En ) =

n

∏Pr ( E ). i

(2.24)

i =1

Example 2.4 Suppose that Vendor 1 provides 40% and Vendor 2 provides 60% of chips used in a computer. It is further known that 2.5% of Vendor 1’s supplies are defective and 1% of Vendor 2’s supplies are defective. What is the probability that a randomly selected chip is both defective and supplied by Vendor 1? What is the same probability for Vendor 2?

Solution: E1 = the event that a chip is from Vendor 1 E2 = the event that a chip is from Vendor 2 D = the event that a chip is defective D|E1 = the event that a defective chip is supplied by Vendor 1 D|E2 = the event that a defective chip is supplied by Vendor 2 Then Pr( E1 ) = 0.40, Pr ( E2 ) = 0.60, Pr ( D | E1 ) = 0.025, and Pr( D | E2 ) = 0.01. From Equation 2.22, the probability that a randomly selected chip is defective and from Vendor 1 is: Pr( E1 ∩ D) = Pr( E1 ) Pr( D | E1 ) = ( 0.4 )( 0.025) = 0.01. Similarly, Pr( E2 ∩ D) = 0.006.

Another probability rule, the inclusion-exclusion principle or addition law of probability, deals with the union of inclusive events. Recall that the union of mutually exclusive events, A ∪ B, is given by one of the axioms of probability to be: Pr ( A ∪ B ) = Pr ( A ) + Pr ( B ) . But what if A and B are not mutually exclusive? The inclusion-exclusion principle applies: Pr ( A ∪ B ) = Pr ( A ) + Pr ( B ) − Pr ( A ∩ B ).

(2.25)

39

Basic Reliability Mathematics

To illustrate the origin of this law, let’s consider the 200 electronic parts that we discussed in Example 2.3. The union of two events L and H includes those parts that do not meet the length requirement, or the height requirement, or both, that is, 48 a total of 23 + 18 + 7 = 48. Thus, Pr ( L ∪ H ) = = 0.24. In other words, 24% of 200 the parts do not meet one or both requirements. We can easily see that Pr ( L ∪ H ) ≠ Pr ( L ) + Pr ( H ), since 0.24 ≠ 0.125+0.15. The reason for this inequality is that the two events L and H are not mutually exclusive. In turn, Pr(L) will include the probability of inclusive events where both L and H fail, (i.e., L ∩ H ). Pr(H) will also include the events where both fail, (L ∩ H ). Thus, joint events are counted twice in the expression Pr ( L ) + Pr ( H ). Therefore, Pr ( L ∩ H ) must be subtracted from this expression. This overlap, which can also be seen in a Venn diagram, leads to the need to correct for this double counting, as shown in Equation 2.25. Since 7 Pr ( L ∩ H ) = = 0.035, then Pr ( L ∪ H ) = 0.125 + 0.15 − 0.035 = 0.24, which is 200 what we expect to get. From Equation 2.25 one can see that if A and B are mutually exclusive, then Pr ( A ∪ B ) = Pr ( A ) + Pr ( B ), again matching the axiom given in Equation 2.14. The addition law of probability for two events can be logically extended to n events: Pr ( E1 ∪ E2 ∪…∪ En ) = Pr ( E1 ) + Pr ( E2 ) + … + Pr ( En ) − Pr  ( E1 ∩ E2 ) + Pr ( E1 ∩ E3 ) + … + Pr ( En −1 ∩ En ) +  Pr ( E1 ∩ E2 ∩ E3 ) + Pr ( E1 ∩ E2 ∩ E4 ) + …… + (−1)n−1 Pr ( E1 ∩ E2 ∩…∩ En ).

(2.26)

Note that Equation 2.25 can be expanded into the form involving the conditional probability by using Equation 2.21: Pr ( A ∪ B ) = Pr ( A ) + Pr ( B ) − Pr ( B A ) Pr ( A ).

(2.27)

Equation 2.26 can be expanded likewise. If all events are independent, then Equation 2.27 simplifies further to Pr ( A ∪ B ) = Pr ( A ) + Pr ( B ) − Pr ( A ) Pr ( B ) ,

(2.28)

which can be reformatted into the more compact form of Pr ( A ∪ B ) = 1 − (1 − Pr ( A )) (1 − Pr ( B )). Generalizing this expression results in the addition law of probability for n independent events: Pr ( E1 ∪ E2 ∪…∪ En ) = 1 −  

n

∏[1 − Pr ( E )]. i

i −1

(2.29)

40

Reliability and Risk Analysis

Example 2.5 A particular type of valve is manufactured by three suppliers. It is known that 5% of valves from Supplier 1, 3% from Supplier 2, and 8% from Supplier 3 are defective. Assume the supplies are independent. If one valve is selected from each supplier, what is the probability that at least one valve is defective?

Solution: D1 = the event that a valve from Supplier 1 is defective D2 = the event that a valve from Supplier 2 is defective D3 = the event that a valve from Supplier 3 is defective D1 ∪ D 2 ∪ D3 is the event that at least one valve from Suppliers 1, 2, or 3 is defective. Since the occurrence of events D1, D2, and D3 is independent, we can use Equation 2.29 to determine the probability of D1 ∪ D 2 ∪ D3. Thus, Pr ( D1 ∪ D2 ∪ D3 ) = 1 − (1 − 0.05)(1 − 0.03)(1 − 0.08 ) = 0.152.

For a large number of events, the addition law of probability can result in combinatorial explosion. Equation 2.26 consists of 2n − 1 terms. A useful method for dealing with this is the rare event approximation. In this approximation, we ignore the joint probability terms (i.e., we treat the n events as if they were mutually exclusive), and use Equation 2.14 (rewritten below as Equation 2.30 for convenience) for the union of events E1 , …, En . This rare event approximation is used if all probabilities of interest, Pr ( Ei ), are small, for example, Pr ( Ei )
x ) =  ∞  f ( t ) dt for a continuous r.v.  x 

∫

The relationship between the ccdf and cdf is shown in Figure 2.6.

(2.50)

48

Reliability and Risk Analysis

F(x)

F(x) ccdf or survival function

cdf

x

x

FIGURE 2.6  The relationship between cdf (left) and ccdf or survival function (right).

A note on terminology: We will use the generic term “probability distribution” to refer to a general distribution. However, we do not recommend the use of the imprecise term “probability distribution” when referring specifically to the pdf or cdf; precision is warranted to avoid confusion of the two. When indicating that a variable takes on a specific distribution form, we use ~ to denote “is distributed,” e.g., X ~ Norm ( µ , σ ). Example 2.10 Let the r.v. T have the pdf  t2  ,  f (T ) =  c  0 



0 < t < 6; otherwise.

Find the value of constant c. Then find Pr(1 0

Range

x≥0

Rate parameter λ representing the constant rate at which x occurs. Typically, a time or distance. (Continued)

64

Reliability and Risk Analysis

TABLE 2.6 (Continued) Exponential Distribution Properties

( )

f x λ = λe

( )

Pdf, f(x) Excel: expon.dist (x, λ , false) Matlab: Exppdf (x, 1/λ ) Python: expon.pdf (x, 0, 1/λ)

−λx

Cdf, F(x)

F x λ = 1− e

Excel: expon.dist (x, λ, true) Matlab: expcdf (x, 1/λ ) Python: expon.cdf (x, 0, 1/λ)

− λt

Properties Mean E(X)

1 λ

Variance Var(X)

1 λ2

Median x 0.5

ln ( 2 ) λ

Skewness Skew(X)

2

Mode

0

Kurtosis Kurt(X)

6

65

Basic Reliability Mathematics

Example 2.17 A system has a constant failure rate of 0.001 failures/hour. What is the probability that the system will fail before t = 1,000 hours? Determine the probability that it works for at least 1,000 hours.

Solution: Calculate the cdf for the exponential distribution at t = 1,000 hours: 1,000

1, 000

Pr(t ≤ 1,000) =

∫ λe

−λ x

dt = −e

0

−λ x

,  0

−1

Pr(t ≤ 1000) = 1 − e = 0.632. Therefore, Pr(t > 1,000) = 1 − Pr(t ≤ 1,000) = 0.368.

2.7.2.2 Continuous Uniform Distribution The continuous uniform distribution is also called the rectangular distribution due to the characteristic shape of its pdf. The uniform distribution is useful for understanding continuous distributions because of its straightforward equations and its mathematical simplicity. It is useful for random number generation, for representing bounds on variables, such as those coming from expert elicitation, and as a non-informative prior in Bayesian analysis. Table 2.7 shows important properties of a uniform distribution. 2.7.2.3 Normal Distribution Perhaps the best known and most important continuous probability distribution is the normal distribution (also called a Gaussian distribution). The parameter μ is the mean of the distribution, and the parameter σ is the standard deviation. We denote a normally distributed random variable as X ~ norm ( µ , σ ). The normal distribution is of particular interest because of the central limit theorem, which states that if we take sufficiently large number (e.g., >30) of independent and identically distributed (i.i.d.) random samples with replacement from a parent population having the mean μ and standard deviation σ, then the distribution of the sample means will be approximately normally distributed. Table 2.8 shows important properties of a normal distribution. Note that the integral in the normal cdf cannot be evaluated in a closed form, so the numerical integration and tabulation of normal cdf are required. However, it would be impractical to provide a separate table for every conceivable value of μ and σ. One way to get around this difficulty is to transform the normal distribution to the standard normal distribution, φ ( z ) , using Equation 2.67, which has a mean of zero (μ = 0) and a standard deviation of 1 (σ = 1). The standard normal distribution has useful properties. Like all pdfs, the area under the curve is one. All three measures of central tendency are the same value: −.5 0. At its mode, f ( z ) = ( 2π ) = 0.40, and F(z) = 0.5. It is symmetric about the y-axis, and thus its skewness is zero. Like most continuous distribution, the normal distribution extends without limit.

66

Reliability and Risk Analysis

TABLE 2.7 Uniform Distribution Properties Parameters & Description X ~ unif ( a, b)

a

a > −∞ b>a a≤x≤b

b Range

Minimum value Maximum value Pdf, f(x)

(

)

1 f x a, b = b−a

Excel: N/A, program directly or see beta distribution Matlab: unipdf (x, a, b) Python: uniform.pdf (x, a, b)

(Continued)

67

Basic Reliability Mathematics

TABLE 2.7 (Continued) Uniform Distribution Properties Cdf, F(x)

(

)

x−a F x a, b = b−a

Excel: N/A, program directly or see beta distribution Matlab: unicdf (x, a, b) Python: uniform.cdf (x, a, b)

Properties Mean E(X) Median x 0.5 Mode

1 (a + b) 2 1 (a + b) 2 Any value in its range

Variance Var(X) Skewness Skew(X) Kurtosis Kurt(X)

1 ( b − a )2 12 0 −

6 5

68

Reliability and Risk Analysis

TABLE 2.8 Normal Distribution Properties µ σ Range

Parameters & Description X ~ norm ( µ , σ )

Location parameter: Mean of the distribution Scale parameter: Standard deviation of the distribution Pdf, f(x)

S=

1  x −µ φ    where: φ is the standard normal pdf. σ  σ 

Excel: norm.dist (x, μ, σ, false) norm.s.dist (z, false) Matlab: normpdf (x, μ, σ) Python: norm.pdf (x, μ, σ)

(Continued)

69

Basic Reliability Mathematics

TABLE 2.8 (Continued) Normal Distribution Properties Cdf, F(x)

(

)

F x µ, σ =

1 σ 2π

x

∫

e

 1  t−µ  2  −       2 σ 

dt =

−∞

Where Φ is the standard normal cdf.

1 1  x−µ + erf = Φ( z ) 2 2  σ 2 

Excel: norm.dist (x, μ, σ, true) norm.s.dist (z, true) Matlab: normcdf (x, μ, σ) Python: norm.cdf (x, μ, σ)

(Continued)

70

Reliability and Risk Analysis

TABLE 2.8 (Continued) Normal Distribution Properties Properties



Mean E(X) Median 0.5 Mode

Variance Var(X) Skewness Skew(X) Kurtosis Kurt(X)









All normal distribution pdfs and cdfs can be transformed into standard form through a simple Z transformation. Appendix A (Table A.1) provides a standard normal cdf table. To use this table, first, transform the r.v. X into Z using Equation 2.67: Z=

X−µ . σ

Therefore, if X takes on value x = x1, Z takes on value z1 = transformation, we can write the equality

(2.67)

( x1 − µ ) . Based on this σ

Pr ( x1 < X < x 2 ) = Pr ( z1 < Z < z2 ).

(2.68)

Example 2.18 A manufacturer states that its light bulbs have a mean life of 1,700 hours and a standard deviation of 280 hours. If the light bulb lifetime, T, is normally distributed, calculate the probability that a randomly light bulb from this population will last less than 1,000 hours.

Solution: First, the corresponding z value is calculated as z=

1,000 − 1,700 = −2.5. 280

Note that the lower tail of a normal pdf goes to −∞, so the formal solution is given by Pr(−∞ < T < 1,000) = Pr (−∞ < Z < −2.5) = 0.0062, which can be represented as Pr(−∞ < T < 1,000) = Pr (−∞ < T ≤ 0) + Pr(0 < T < 1,000). The first term on the right side does not have a physical meaning since time cannot be negative. A more proper distribution, therefore, would be a truncated normal

71

Basic Reliability Mathematics distribution; such a distribution will only exist in t ≥ 0. Truncation will be covered later in this chapter. This is unnecessary for this problem because 1,700   −10 Pr(−∞ < T < 0) = Pr (−∞ < Z ) = Pr  −∞ < Z < −  = 6.34 × 10 ,  280  which can be considered as negligible. So, finally, one can write Pr(−∞ < T < 1,000) ≈ Pr (0 < T < 1,000) ≈ Pr (−∞ < Z < −2.5) = 0.0062.

2.7.2.4 Lognormal Distribution The lognormal distribution is closely related to the normal distribution. If r.v. X is lognormally distributed with parameters µ and σ , then the r.v. ln(X) is normally distributed with a mean of µ and a standard deviation of σ . That is, if X ~ lognorm ( µ , σ ) then ln ( X ) ~ norm ( µ , σ ) . Cautions: the parameters of the lognormal distribution are not the mean and variance of the lognormal distribution (or r.v. X); rather, they are the mean and variance of the normally distributed ln(X). This distribution is often used to model an r.v. that can vary by several orders of magnitude. One major application of the lognormal distribution is to model an r.v. that results from multiplication of many independent variables. Table 2.9 shows important properties of the lognormal distribution. Some applications of the lognormal distribution parameterize it in terms of its median and an error factor, EF, EF =

95th percentile 95th percentile 50th percentile = = . 5th percentile 50th percentile 5th percentile

(2.69)

Like the normal distribution, the lognormal cdf does not have a closed form solution. It can be solved by transforming into Z space and using the standard normal distribution. Transforming the lognormal distribution into Z space using a slightly different Z: Z=

ln x − µ . σ

(2.70)

Given the mean and variance of X, E(X) and var(X), respectively, you can transform those into the parameters of the lognormal distribution as such:   E(X ) µ = ln   Var ( X )2  1+  E ( X )2

   ,    

(2.71)

 Var ( X )2  σ = ln  1 + . E ( X )2  

(2.72)

and

72

Reliability and Risk Analysis

TABLE 2.9 Lognormal Distribution Properties Parameters & Description X ~ lognorm ( µ , σ )

µ σ Range

−∞ < µ < ∞ σ > 0 x>0

Scale parameter Shape parameter Pdf, f(x)

(

)

f x µ, σ =

1 xσ 2π

e

1  ln x − µ  2 −   2 σ 

=

1  ln x − µ  φ  xσ  σ 

where φ ( z ) is the standard normal pdf

Excel: lognorm.dist (x, µ , σ , false) Matlab: lognpdf (x, µ , σ ) Pythona: lognorm.pdf (x, σ, 0, np.exp(μ))

(Continued)

73

Basic Reliability Mathematics

TABLE 2.9 (Continued) Lognormal Distribution Properties Cdf, F(x)

(

)

F x µ, σ = =

1 σ 2π

x

∫

−∞

1 e x

 1  ln t − µ  2  −    σ    2 

dt

1 1  ln x − µ   ln x − µ  , + erf =φ  σ  2 2  σ 2 

Excel: lognorm.dist (x, µ , σ , true) Matlab: logncdf (x, µ , σ ) Python: lognorm.cdf (x, σ, 0, np.exp(μ))

where ( z ) is the standard normal cdf.

(Continued)

74

Reliability and Risk Analysis

TABLE 2.9 (Continued) Lognormal Distribution Properties Properties Mean E(X) Median 0.5 Mode

a

e

 σ2  µ−  2  

eµ e

( µ −σ 2 )

Variance Var(X)

 (σ 2 )  ( 2 µ +σ 2 ) e − 1 ⋅ e  

Skewness Skew(X) Kurtosis Kurt(X)

 (σ 2 )  (σ 2 −1)  e + 2 e

e

(4σ 2 ) + 2e(3σ 2 ) + 3e(2σ 2 ) − 6

Python’s SciPy package requires the “standardized” form of the lognormal distribution; using “np. exp(μ)” translates the parameterization given here for use in Python.

Example 2.19 The lifetime of a motor has a lognormal distribution. What is the probability that the motor lifetime is at least 10,000 hours if the mean and variance of the normal random variable are 10 hours and 1.3 hours? Then, find the mean and variance of motor failure times.

Solution: The corresponding z value is calculated as: z=

ln (10, 000 ) − 10 = −0.607. 1.3

Looking up this value in Appendix A and interpolating gives: Pr ( X ≤ 10, 000 ) = 0.271. So Pr( X > 10,000) = 1 − 0.271 = 0.728. To solve for the mean of the lognormal distribution:  σ2   µ+  2 

E ( X ) =  e

 1.32   10+  2 

= e

= 51, 277.12 hours.

Then, solving for the standard deviation:  (σ 2 )  ( 2 µ + σ 2 ) Stdev ( X ) = √ Var ( X ) =  e − 1 e   =

(e

(1.32 )

)

−1 ⋅e

(2⋅10 +1.32 ) = 107,797 hours.

2.7.2.5 Weibull Distribution The Weibull is another distribution in the exponential family. It is widely used to represent the time to failure or life duration of components as well as systems.

75

Basic Reliability Mathematics

The parameter β is a shape parameter and α is a scale parameter, that is, X ~ weibull (α ,  β ) . Note that the two-parameter Weibull distribution has multiple parameterizations that are used in reliability. Additionally, there is a three-parameter version of the Weibull distribution which is not covered in this book. Several properties of the Weibull distribution rely on a function known as the gamma function Γ ( x ) . See Section 2.7.2.6 for explanation of this function. When β = 1, the Weibull distribution is reduced to the exponential distribution with λ = 1/α, so the exponential distribution is a particular case of the Weibull distribution. For the values of β > 1, the distribution becomes bell-shaped with some skew. We will elaborate further on this distribution and its use in reliability analysis in Chapter 3. Table 2.10 shows important properties of the Weibull distribution. 2.7.2.6 Gamma Distribution (and Chi-Squared) The gamma distribution can be thought of as a generalization of the exponential distribution. For example, if Ti , the time between successive failures of a system has α

an exponential distribution with (parameter λ ), then an r.v. T such that T =

∑T foli

i =1

lows a gamma distribution with parameters α and β = 1/λ . In the given context, T represents the cumulative time to the α th failure. A different way to interpret this distribution is to consider a situation in which a system is subjected to shocks occurring according to the Poisson process (with parameter λ = 1/β). If the system fails after receiving α shocks, then the time to failure of such a system follows a gamma distribution. A simple example of this system is a redundant system with α standby parallel components. Failure of each standby component can be viewed as a shock. The system fails after the αth component fails (i.e., the last shock occurs). Standby systems will be discussed later in Chapter 6. If α = 1, the gamma distribution is reduced to an exponential distribution. Another important special case of the gamma distribution is the case when β = 2 and α = df/2, where df is a positive integer, called the number of degrees of freedom. This oneparameter distribution is known as the chi-squared (χ2) distribution. This distribution is widely used in reliability data analysis with some applications discussed in Chapter 4. Table 2.11 shows important properties of the gamma distribution. The gamma distribution uses a function Γ ( n ) called the gamma function.2 The gamma function is: ∞

∫

Γ ( n ) = x n −1e − x dx.

(2.73)

0

For integer values of n, Γ(n)=(n-1)! where ! denotes factorial 2.7.2.7 Beta Distribution The beta distribution is a useful model for r.v.s that are distributed in a finite interval. The pdf of the standard beta distribution is defined over the interval [0,1], making it particularly useful for Bayesian analysis. Like the gamma distribution, the cdf of 2

Caution: do not get the gamma functions confused with the gamma distribution; we recommend being precise in using the term function vs. distribution to avoid the confusion.

76

Reliability and Risk Analysis

TABLE 2.10 Weibull Distribution Properties Parameters & Description X ~ weibull (α ,  β )

α



β Range

Scale parameter Shape parameter Typically, a time or distance Pdf, f(x)

(

)

f x α,β =

β −1

βx e αβ

β  x −  α

Excel: weibull.dist (x, β , α , false) Matlab: wblpdf (x, α , β ) Python: weibull_min.pdf (x, β, 0, α)

(Continued)

77

Basic Reliability Mathematics

TABLE 2.10 (Continued) Weibull Distribution Properties Cdf, F(x)

(

)

F x α,β = 1− e

β  x −  α

Excel: weibull.dist (x, β, α, true) Matlab: wblcdf (x, α , β ) Python: weibull_min.cdf (x, β, 0, α)

(Continued)

78

Reliability and Risk Analysis

TABLE 2.10 (Continued) Weibull Distribution Properties Propertiesa Mean E(X)

 1 αΓ  1 +  β 

Variance Var(X)

2   1   2   α 2 Γ  1 +  −   1 +    β   β      

Median x 0.5

α ( ln ( 2 ))

Skewness Skew(X)

 3 α 3 Γ  1 +  − 3µσ 2 − µ 3 β  σ3

Kurtosis Kurt(X)

−  6Γ14 + 12Γ12  Γ 2 − 3Γ 22 − 4 Γ1  Γ 3 + Γ 4

1/ β

 β − 1 α  β   

Mode

a

1/ β  

if β ≥ 1

(Γ

2

− Γ12

)

2

Otherwise, no mode exists

Γi refers to the gamma function (Equation 2.73) evaluated at i.

TABLE 2.11 Gamma Distribution Properties Parameters & Description X ~ gamma (α ,  β ) α β Range

Shape parameter Scale parameter Typically, a time or distance



Pdf, f(x) x  − 1  x α −1e β  for continuous α α  β Γ (α ) f x α,β =  x − 1  α −1 β  β α (α − 1)! x e  for integer α 

(

)

 

Excel: gamma.dist (x, α, β, false) Matlab: gampdf (x, α, β) Python: gamma.pdf (x, α, 0, 1/β)a

(Continued)

79

Basic Reliability Mathematics

TABLE 2.11 (Continued) Gamma Distribution Properties Pdf, f(x)

Cdf, F(x) x  x − 1  x α −1e β   dx   for continuous α α  β Γ (α )  0 F x α,β =  x α −1 −  xn  for integer α 1− e β  n β n!  n= 0

(

)

∫

∑

Excel: gamma.dist (x, α, β, true) Matlab: gamcdf (x, α, β) Python: gamma.cdf (x, α, 0, 1/β)

Cannot be written in closed form; often written as  x  x 1 γ α, F ( x α ,  β ) = where γ  α ,  is known as the lower Γ (α )  β   β incomplete gamma function.

(Continued)

80

Reliability and Risk Analysis

TABLE 2.11 (Continued) Gamma Distribution Properties Cdf, F(x)

Properties Mean E(X) Median x 0.5 Mode

αβ Must be solved numerically

(α − 1) β if α ≥ 1 Otherwise, no mode exists

a

Variance Var(X) Skewness Skew(X) Kurtosis Kurt(X)

αβ 2 2 α 6 α

Python’s SciPy package uses the single-variable parameterization of the Gamma distribution. Using “scale=1/β” translates the parameterization given here for use in Python.

the beta distribution cannot be written in closed form. It is expressed in terms of the incomplete beta function, It(α, β). For the special case of α = β = 1, the beta distribution reduces to the standard uniform distribution. Practically, the distribution is not used as a time to failure distribution. However, the beta distribution is widely used as an auxiliary distribution in nonparametric classical statistical distribution estimations, and as a prior distribution in the Bayesian statistical inference, especially when the r.v. can only range between 0 and 1: for example, in a reliability test or any other probability estimate. Table 2.12 shows important properties of the beta distribution.

2.7.3

truncated distributions

Truncation is needed when the random variable cannot exist in certain extreme ranges. An example of truncation is when the length of a product is normally

81

Basic Reliability Mathematics

TABLE 2.12 Beta Distribution Properties Parameters & Description X ~ beta (α, β)

α >0 β >0 0 ≤ x ≤1

α β Range

Shape parameter Shape parameter Pdf, f(x)

Γ (α + β ) α −1 β −1 f x α,β = x (1 − x ) Γ (α ) Γ ( β )

(

)

Where the beta function is B (α ,  β ) =

Γ (α ) Γ ( β ) Γ (α + β )

Excel: beta.dist (x, α, β, false) Matlab: betapdf (x, α, β) Python: beta.pdf (x, α, β, 0, 1)

(Continued)

82

Reliability and Risk Analysis

TABLE 2.12 (Continued) Beta Distribution Properties Cdf, F(x) F ( x;α , β ) =

Γ (α , β ) Γ (α + β ) B x α,β x a−1 (1− x )β −1 dx = Γ (α ) Γ ( β ) Γ (α ) Γ ( β ) x

(

∫ 0

(

)

)

where the integral above is B x α , β , the incomplete beta function. The cdf can also be written as I x (α ,  β ) , the regularized incomplete beta function:

(

Excel: beta.dist (x, α, β, true) Matlab: betacdf (x, α, β) Python: beta.cdf (x, α, β, 0, 1)

B x α,β ) B( (α , β ) ) .

Ix x α , β =

B(α, β) is the marginalized version of the incomplete beta function.

(Continued)

83

Basic Reliability Mathematics

TABLE 2.12 (Continued) Beta Distribution Properties Properties Mean E(X) Median x 0.5

Mode

α α +β

≈

α−

Variance Var(X) 1 3

α +β−

2 3

 for α ,  β > 1 

α −1 for α, β > 1 α +β−2

αβ

(α + β )2 (α + β + 1)

Skewness Skew(X)

2 (β − α ) α + β + 1

Kurtosis Kurt(X)

2 6 (α − β ) (α + β + 1) − αβ (α + β + 2 )    αβ (α + β + 2 ) (α + β + 3)

See references for other combinations.

(α + β + 2)

αβ

distributed about its mean. The normal distribution has a range from −∞ to +∞, however it is obviously not realistic for a physical dimension like length or time to be negative. Therefore, a truncated form of the normal distribution needs to be used if the normal distribution is used to model the length of a product. Another example of truncation is when the existence of a defect is unknown due to the defect’s size being less than the inspection threshold. A truncated distribution is therefore a conditional distribution that restricts the domain of another probability distribution. The following general equation forms apply to the truncated distribution, where f0 ( x ) and F0 ( x ) are the pdf and cdf of the non-truncated distribution. The pdf of a truncated distribution can be expressed by  f0 ( x )   a < x ≤ b,   F b ( f ( x ) =  0 ) − F0 ( a )  0 otherwise. 

(2.74)

As such the cdf of a truncated distribution would be  0  x ≤ a,   x  f0 ( t ) dt  a   a < x ≤ b,   F (x) =   F0 ( b ) − F0 ( a )  1  x > b.   

∫

2.7.4

(2.75)

Multivariate distributions

Thus far, we have discussed probability distributions that describe a single random variable and one-dimensional sample spaces. There exist, however, situations in which more than one possibly dependent r.v. is simultaneously measured

84

Reliability and Risk Analysis

and recorded. For example, in a study of human reliability in a control room situation, one can simultaneously estimate the r.v. T representing time that various operators spend to fulfill an emergency action, and the r.v. E representing the level of training the operators have had for performing these emergency actions. Since one expects E and T to have some relationship (e.g., better-trained operators act faster than less-trained ones), a joint distribution of both r.v.s T and E can be used to express their mutual dispersion. Let X and Y be two r.v.s (not necessarily independent). The joint pdf (or bivariate pdf) of X and Y is denoted by f(x, y). If X and Y are discrete, the joint pdf can be denoted by Pr(X = x, Y = y), or simply Pr(x, y), just as we described when we introduced the concept of joint probability. Thus, Pr(x, y) gives the probability that the outcomes x and y occur simultaneously. For example, if r.v. X represents the number of electric circuits of a given type in a process plant and Y represents the number of failures of the circuit type in the most recent year, then Pr(7, 1) is the probability that a randomly selected process plant has seven circuits and that one of them has failed once in the most recent year. The function f(x, y) is a joint pdf of continuous r.v.s X and Y if 1.

f ( x ,  y )   ≥   0, − ∞ 0, Pr ( X = x ) Pr ( x )

Pr(Y = y X = x ) =

(2.86)

where X and Y are discrete r.v.s. Similarly, one can extend the same concept to continuous r.v.s X and Y and write f x y =

( )

f ( x ,  y ) ,  fY ( y ) > 0,  fY ( y )

(2.87)

( )

f ( x ,  y ) ,  f X ( x ) > 0, fX ( x )

(2.88)

or f y x =

where Equations 2.86–2.88 are called the conditional pdfs of discrete and continuous r.v.s, respectively. The conditional pdfs have the same properties as any other pdf. Similar to the discussions earlier in this chapter, if X and Y are independent, then f ( x y) = f ( x ). This would lead to the conclusion that for independent r.v.s X and Y, f ( x ,  y ) = f X ( x ) ⋅ fY ( y ) , 

(2.89)

if X and Y are continuous, and Pr ( x ,  y ) = Pr ( x ) ⋅ Pr ( y ) ,

(2.90)

86

Reliability and Risk Analysis

if X and Y are discrete. Equation 2.89 and 2.90 can be expanded to a more general case as f ( x1 ,  …,  x n ) = f1 ( x1 ) ⋅…⋅ fn ( x n ) , 

(2.91)

where f ( x1 , …,  x n ) is a joint pdf of X1 , …, X n and f1 ( x1 ) ,  f2 ( x 2 ) , … fn ( x n ) are marginal pdfs of X1 , X 2 , …,  X n respectively. For more general cases where X1 , …, X n are not independent f ( x1 ,  x 2 , …, x n ) = f ( x1 x 2 , …, x n ) f ( x 2 , …,  x n ) ,  = f ( x1 ) ⋅ f ( x 2 x1 ) ⋅…⋅ f ( x n −1 x1 ,  …, x n − 2 ) ⋅ f ( x n x1 ,  …, x n ) . (2.92) The marginal pdf of the general joint continuous and discrete joint pdf can be, respectively, expressed as,

∫ ∫ ∫ f ( x ,  x , …, x ) dx dx …dx ,

(2.93)

∑…∑∑Pr(x ,  x , …,  x ).

(2.94)

f ( x1 ) = … xn

1

2

2

n

3

n

x3 x 2

Pr ( x1 ) =

1

x2

x3

2

n

xn

And the conditional pdfs are: f ( x1 x 2 , …,  x n ) =

Pr ( x1 x 2 , …,  x n ) =

f ( x1 , x 2 , …,  x n ) = f ( x 2 , …, x n )

∫

Pr ( x1 ,  x 2 , …, x n ) = Pr ( x 2 , …,  x n )

f ( x1 , x 2 , …,  x n ) x1

f ( x1 , x 2 , …,  x n ) dx1

, 

(2.95)

.

(2.96)

Pr ( x1 ,  x 2 , …, x n )

∑

x1

Pr( x1 ,  x 2 , …,  x n )

Example 2.20 Let T1 represent the time (in minutes) that a machinery operator spends to locate and correct a routine problem and let T2 represent the length of time (in minutes) that the operator needs to spend reading procedures for correcting the problem. If T1 and T2 are represented by the joint probability distribution,

(

)

 c t 1/ 3 + t 1/ 5 ,    1 2 f ( t1 ,  t2 ) =  0 

60 > t1 > 0, 10 > t2 > 0; otherwise.

87

Basic Reliability Mathematics Find a. The value of constant c. b. The probability that an operator can correct the problem in less than 10 minutes. Assume that the operator in this accident should spend less than 2 minutes to read the necessary procedures. c. Whether r.v.s T1 and T2 are independent.

Solution: t1 = 60 t2 = 10

∫ ∫

a. Pr(t1 < 60, t2 < 10) =

0

0

 1 1 c  t13 + t25  dt1dt2 = 1,  

c = 3.92 × 10 −4. t1 = 10 t2 = 2

b. Pr(t1 < 10, t2 < 2) =

∫ ∫ 3.92 × 10 (t −4

0

1/ 3 1

)

+ t21/ 5 dt1dt2 = 0.02.

0

t1 = 60

c. f ( t2 ) =

∫ f (t , t ) dt = 3.92 × 10 (176.17 + 60t ). 1

2

1

−4

1/ 5 2

0

t2 = 10

Similarly, f ( t1 ) =

∫

(

)

f ( t1 ,  t2 ) dt2 = 3.92 × 10 −4 10t21/ 3 + 13.21 .

0

Since f ( t1 ,  t2 ) ≠   f ( t1 ) ⋅ f ( t2 ), T1 and T2 are not independent.

2.8 2.1

EXERCISES Simplify the following Boolean functions

( A ∩ B ∪ C ) ∩ B. b. ( A ∪ B ) ∩ ( A ∪ B ∩ A ) . a. c.

A ∩ B ∩ B ∩ C ∩ B.

(

(

)

)

2.2

Reduce the Boolean function A ∩ B ∩ C ∪ C ∪ A ∪ B .

2.3

Simplify the following Boolean expression ( A ∪ B ) ∩ A  ∪ B ∩ A .

2.4

Reduce the Boolean function G = ( A ∪ B ∪ C ) ∩ A ∩ B ∩ C ∩ C. If

2.5

Pr ( A ) = Pr ( B ) = Pr (C ) = 0.9, what is Pr(G)? Simplify the following Boolean equations

2.6

(

(

(

)

a.

( A ∪ B ∪ C ) ∩ A ∩ B ∩ C ∩ C.

b.

( A ∪ B ) ∩ B.

(

) (

)

Reduce the Boolean equation A ∪ ( B ∩ C ) ∩ B ∪ ( D ∩ A ) .

)

)

88

Reliability and Risk Analysis



 

   



 



 

Basic Reliability Mathematics



89

90

Reliability and Risk Analysis





     f x a,  b,  c =      

(

)

2( x − a) ( b − a )( c − a ) 2 b−a 2( b − x ) ( b − a )( b − c ) 0

a ≤ x ≤ c,  x = c,  c ≤ x ≤ b,  otherwise.

where, • Parameters a −∞ ≤ a < b Minimum Value. a is the lower bound, • Parameter b, a < b < ∞ Maximum Value. b is the upper bound, • Parameter c a ≤ c ≤ b Mode Value. c is the mode of the distribution (top of the triangle).

Basic Reliability Mathematics

91

A large electronics store is returning a set of failed televisions and you need to know how many shipping containers to prepare. Assume however that all you know is that the minimum number of failed TV sets is 40, the maximum number that failed is 95, and the mode value of TV failures is 60. Draw what the triangular distribution of this failure model would look like. What is the probability that between 40 and 65 TVs are failed? 2.30 Derive the cdf for the triangular distribution given in Problem 2.29. Plot the cdf and pdf if a = 15, b = 25, c = 50.

REFERENCES Cox, R. T, “Probability, Frequency and Reasonable Expectation.” American Journal of Physics, 14(1), 1–13, 1946. Hahn, G. J., and S. S. Shapiro, Statistical Models in Engineering. Wiley, New York, 1994 Johnson, N. L., S. T. Kotz, and N. Balakrishnan, Continuous Univariate Distributions, Volume 2. John Wiley & Sons, New York, 1995. Nelson, W. B., Applied Life Data Analysis. John Wiley & Sons, New York, 2003. O’Connor, A. N., M. Modarres, and A. Mosleh, Probability Distributions Used in Reliability Engineering, DML International, Washington, DC 2019.

3

Elements of Component Reliability

In this chapter, we discuss the basic elements of component reliability estimation. We start with a mathematical definition of reliability and define commonly used terms and metrics. These definitions also apply to physical items of various complexity (e.g., components, subsystems, systems). We then focus on the use of probability distributions in component reliability analysis in the rest of the chapter.

3.1 DEFINITIONS FOR RELIABILITY Reliability has many connotations. In general, it refers to an item’s ability to successfully perform an intended function during a mission. The longer the item performs its intended function without failure, the more reliable it is. There are both robustness in engineering design and probabilistic views of reliability. The robustness view deals with those engineering design and analysis activities that extend an item’s life by controlling or eliminating its potential failure modes. Examples include designing stronger and more durable parts, reducing harmful environmental conditions, minimizing loads and stresses applied to an item during its use, and providing a condition monitoring, predictive or preventive maintenance program to minimize the occurrence of failures. The probabilistic view measures the reliability of an item in terms of the probability of the successful achievement of the item’s intended function. The probabilistic definition of reliability, given in Section 1.4, is the mathematical representation of this viewpoint. The right-hand side of Equation 1.1 denotes the probability that the r.v. Tfail , representing time to failure, exceeds a specified mission time or time of interest tinterest under specified operating stress conditions c1 , …, cn . Other representations of the r.v. Tfail include the number of cycles to failure, distance to failure such as kilometers to failure that applies to vehicles, and so on. In the remainder of this book, we consider primarily the time to failure representation, although the same mathematical formulations and treatment equally apply to other representations. Expected design and operating conditions c1 , …, cn are often implicitly considered and not explicitly shown; therefore, Equation 1.1 is written in a simplified form of Equation 1.2. We use Equation 1.2 in the remainder of this chapter. However, the conditions should be explicitly considered when stresses and operating conditions are relevant, such as in accelerated testing and causal modeling.

3.1.1

reliability function

Let us start with the formal definition of reliability given by Equation 1.1. Let f(t) denote a pdf representing the item failure time r.v. T fail . According to Equation 2.41, the probability of failure of the item as a function of time is defined by DOI: 10.1201/9781003307495-3

93

94

Reliability and Risk Analysis

Pr ( T fail ≤ t ) =

t

∫ f (θ ) dθ = F (t )  for t ≥ 0,

(3.1)

0

where F(t) denotes the probability that the item will fail sometime up to the time of interest t. Equation 3.1 is the unreliability of the item. Formally, we can call F(t) (which is the time to failure cdf) the unreliability function. Conversely, we can define the reliability function (or the survivor or survivorship function) as ∞

R (t ) = 1 − F (t ) =

∫ f ( x ) dx.

(3.2)

t

Let R(t) be the reliability function of an item at time t. The probability that the item will survive for (additional) time x, given that it has survived for time t, is called the conditional reliability function, and is given by R ( t + x|t ) =

R (t + x ) . R (t )

(3.3)

Some literature defines x as a new random variable defined as the time after reference point t. Typically this framing is used with t = t0  being a warranty or burn-in time. In R( x ) that framing, x = 0 at t0 . In this case, Equation 3.3 becomes: R( x | t0 ) = . R ( t0 ) For systems where their failure depends only on the number of times the item is used (rather than length of time) such as some demand items, the cycle-based reliability model may be used. Let N be the integer r.v. representing the cycle of use, then the pmf, cdf, and the reliability function at the point of n uses would be expressed, respectively, by f ( n ) = Pr ( N = n ) , 

(3.4)

n

F ( n ) = Pr ( N ≤ n ) =

∑ f (i ) , 

(3.5)

i=1

n

R(n) = 1 − F (n) = 1 −

∑ f (i ).

(3.6)

i=1

Similarly, the conditional reliability if the item has survived n uses is R ( n + n′ n ) =

R ( n + n′ ) .  R(n)

(3.7)

95

Elements of Component Reliability

Recall that integration is the continuous analog of summation, it is clear to see that Equations 3.1–3.3 apply to continuous variables whereas Equations 3.4–3.7 apply only to discrete variables.

3.1.2

Mttf, Mrl, Mtbf, and quantiles

The basic characteristics of time to failure distribution and basic reliability measures can be expressed in terms of the pdf, f(t), cdf, F(t), or reliability function, R(t). We can also describe reliability in terms of several types of characteristics, such as the central tendency. The mean time to failure (MTTF), for example, illustrates the expected time during which a nonrepairable item will perform its function successfully (sometimes called expected life). According to Equation 2.54, ∞

∫

MTTF = E ( t ) = tf ( t ) dt. 

(3.8)

0

If lim tf ( t ) = 0, integrating by parts reduces Equation 3.8 to t→∞

∞

∫

E ( t ) = R ( t ) dt. 

(3.9)

0

It is important to make a distinction between MTTF and the mean time between failures (MTBF). The MTTF is associated with nonrepairable components (i.e., replaceable components), whereas the MTBF is related to repairable components only. For MTBF, the pdf in Equation 3.8 can be the pdf of time between the first failure and the second failure, the second failure and the third failure, and so on for each repairable item. If we have surveillance, and the item is completely renewed through replacement, maintenance, or repair without delay, the MTTF coincides with MTBF. Theoretically, this means that the renewal process is assumed to be perfect. That is, the item that goes through repair or maintenance is assumed to exhibit the characteristics of a new item. In practice this may not be true. In this case, one needs to determine the MTBF for the item for each renewal cycle (each ith interval of time between two successive failures). However, the approach based on the as good as new assumption can be adequate for some reliability considerations. The mathematical aspects of MTBF will be discussed in Chapter 7. The mean residual life (MRL) at time t is defined as the expected remaining life given the component has survived to time, t. ∞

∫

MRL  ( t ) = R ( x t ) dx = 0

where t′ = x + t.

∞

1 R ( t ′ ) dt ′.  R (t )

∫ t

(3.10)

96

Reliability and Risk Analysis

For a continuous r.v. T with cdf F(t), the p-level quantile denoted as tp, is expressed as F ( t p ) = p, 0 < p < 1. The median is defined as the quantile of the level of  p = 0.5. A quantile is often called the “100p% point,” or the “100pth percentile.” In reliability studies, the 100pth percentile of time to failure is the point at which the probability of an item’s failure is p. For example, the B10 life is the time at which 10% of the components are expected to fail. The most popular percentiles used in reliability are the first, fifth, tenth, and 50th percentiles. Example 3.1 A device time to failure follows the exponential distribution. If the device has survived up to time t, determine its MRL.

Solution: According to Equation 3.10,

∫ MRL ( t ) =  

∞ 0

τ f ( t + τ ) dτ

∫

∞ t

f (τ ) dτ

∫ =

∞ 0

τλ e − λ(t +τ ) dτ

∫

∞ t

λ e − λτ dτ

e λt =

∞

∫ τλe 0

e

− λt

− λτ

dτ

=

1 . λ

This underlines a pitfall of the exponential distribution as the conditional MRL remains the same and independent of time, t, as the unconditional MTTF. This notion will be further discussed in Section 3.2.1.

3.1.3

hazard rate and failure rate

It is often useful to know the rate of failure at a particular time in an item’s lifetime. The hazard rate, h(t), is the instantaneous failure rate of an item of age t. The hazard rate is the conditional probability that an item which has survived up to time t will fail during the following small interval of time ∆t, as ∆t approaches zero. h ( t ) = lim

∆t → 0

1 F ( t + ∆t ) − F ( t ) f ( t ) = ,  ∆t R (t ) R (t )

(3.11)

From Equation 3.11 it is evident that h(t) is also a time to failure conditional pdf similar to Equation 3.3. The hazard rate is sometimes called the failure rate. In this book, we use the term failure rate only in the context of the constant failure rate. The hazard rate can also be expressed in terms of the reliability function as h (t ) = −

d  ln R ( t )  ,  dt 

(3.12)

so that t

R (t ) = e

∫

− h ( x ) dx 0

. 

(3.13)

97

Elements of Component Reliability

The integral of the hazard rate in the exponent is called the cumulative hazard function, H(t): t

∫

H ( t ) = h ( x ) dx.

(3.14)

0

The expected value associated with the conditional pdf represented by the hazard rate is called the residual MTTF or simply MRL. For a time interval, t, the mean hazard rate is given by E [ h ( t )] =

t

1 h ( x ) dx, t

∫

(3.15)

0

therefore, E [ h ( t )] = −

lnR ( t ) , t

(3.16)

H (t ) . t

(3.17)

and E [ h ( t )] =

If we want to find the mean hazard rate over a life percentile, t p, then F ( t p ) = 1 − R ( t p ) = p, and E  h ( t p )  = −

ln (1 − p ) . tp

(3.18)

The hazard rate is an important function in reliability analysis since it shows changes in the probability of failure over the lifetime of a component. In practice, h(t) often exhibits a bathtub shape and is called a bathtub curve. An example of a bathtub curve is shown in Figure 3.1. h(t)

Infant Mortality

Chance Failures

FIGURE 3.1 Example bathtub curve.

Wear-out t

98

Reliability and Risk Analysis

Generally, a bathtub curve can be divided into three regions. The burn-in or early failure region exhibits a decreasing hazard rate (sometimes called decreasing (instantaneous) failure rate, DFR), characterized by early failures attributable to defects or poor quality control in materials, manufacturing, or construction of the components. Most components do not experience this early failure characteristic, so this part of the curve represents the population and not individual units. Often, warranties are designed to address concerns about early failures. The chance failure region of the bathtub curve exhibits a reasonably constant hazard rate (or constant failure rate, CFR), characterized by random failures of the component. In this period, many mechanisms of failure due to chance events, environmental conditions, or due to complex underlying physical, chemical, or nuclear phenomena give rise to this approximately constant failure rate. The third region, called the wear-out region, which exhibits an increasing hazard rate (or increasing instantaneous failure rate, IFR), is characterized mainly by complex aging phenomena of an item approaching its end of useful life. Here, the component deteriorates (e.g., due to accumulated fatigue, corrosion, or other types of damage) and is more vulnerable to outside shocks. It is helpful to note that these three regions can be different for different types of components. There are different time to failure distributions that can properly model increasing, constant, or decreasing hazard rate, which we will come back to in Section 3.2. There are drastic differences between the bathtub curves of various components. Figures 3.2 and 3.3 show typical bathtub curves for mechanical and electrical devices, respectively. These figures demonstrate that electrical items generally exhibit a relatively long chance failure period, followed by more abrupt wear-out, whereas mechanical items tend to wear-out over a longer period of time. Figure 3.4 shows the effect of various levels of stress on a component. That is, when under higher stress operation, the component will have shorter burn-in and chance failure regions and higher overall hazard rate values. As stress level increases, the chance failure region decreases and premature wear-out occurs. Therefore, it is important to minimize stress factors, such as a harsh operating environment, to maximize reliability. For the cycle-based (discrete r.v.) items the corresponding hazard rate is defined by

FIGURE 3.2

h (n) =

Pr ( N = n ) f (n) R ( n − 1) − R ( n ) = = ,  Pr ( N ≥ n ) R ( n − 1) R ( n − 1)

Typical bathtub curve for mechanical devices.

(3.19)

99

Elements of Component Reliability

FIGURE 3.3

Typical bathtub curve for electrical devices.

FIGURE 3.4 Effect of various stress levels on components.

where, f(n) is the pmf in Equation 3.4 and R ( n − 1) is the reliability function in Equation 3.6 up to but excluding cycle n. Note that in the cycle-based reliability the cumulative hazard rate H ( n ) = − ln [ R ( n )] ≠

n

∑h (i ). In a discrete counterpart of the i=0

continuous distribution the values of the hazard rate do not converge. To overcome this limitation, a function that maintains the monotonicity property called the second failure rate was proposed by Gupta et al. (1997) as r ( n ) = ln

R ( n − 1) = − ln [1 − h ( n )]. R(n)

Example 3.2 The hazard rate h(t) of a device is approximated by  0.1 − 0.001t ,   h (t ) =  −0.1 + 0.001t ,   

0 ≤ t ≤ 100, t > 100,

(3.20)

100

Reliability and Risk Analysis

as shown in the figure below. Find the pdf and the reliability function for t ≤ 200.

Solution: For 0 ≤ t ≤100, t

∫

t

h ( x ) dx =

0

∫ 0

( 0.1 − 0.001x ) dx =  0.1x −

t

0.001 2  x  = 0.1t − 0.0005t 2 , 0 2

thus R ( t ) = e −0.1t + 0.0005t . 2

Using Equation 3.4, one obtains f ( t ) = ( 0.1 − 0.001t ) e

(−0.1t + 0.0005t 2 ) .

Note that R (100 ) = e −5 = 0.0067, For t > 100, h ( t ) = −0.1 + 0.001t. Accordingly, t

∫ ( 0.1− 0.001x )dx

R ( t ) = R (100 ) ⋅ e100 = 0.0067e

(0.1t − 0.0005t 2 −5) , t

f ( t ) = 0.0067 ( −0.1 + 0.001t ) e

∫ ( 0.1− 0.001x )dx

100

= 0.0067 ( −0.1 + 0.001t ) e

(0.1t − 0.005t 2 −5) .

101

Elements of Component Reliability

3.2

COMMON DISTRIBUTIONS IN COMPONENT RELIABILITY

This section discusses exponential, normal, lognormal, Weibull, gamma, and beta distributions commonly used as time to failure distribution models for components. Characteristics of these distributions were discussed in Chapter 2. Their hazard rates are discussed in this section, along with usage in component reliability analysis. Table 3.1 lists the hazard rate functions for continuous reliability models and summarizes applications for some important distributions used in reliability. Discrete reliability models are more involved, and the readers are referred to Bracquemond and Gaudoin (2003).

3.2.1

exPonential distribution and Poisson distribution

The exponential distribution is one of the most used distributions in reliability analysis. This can be attributed primarily to its simplicity and to the fact that it gives a constant hazard rate model. That is, for the exponential distribution, h ( t ) = λ . In the bathtub curve, this distribution only represents the chance failure region. It is evident that this distribution might be adequate for components whose chance failure region is long in comparison with the other two regions. A constant failure rate is applicable for some electronic components and mechanical components, especially in certain applications when new components are screened out (e.g., during quality control) and only those that are determined to have passed the burn-in period are used. Furthermore, the exponential distribution is a possible model for representing more complex items, systems, and nonredundant components consisting of many interacting parts. In Section 2.7.2.1, we noted that the exponential distribution can be introduced using the HPP. Now let us assume that each failure in this process is caused by a random shock, and the number of shocks occurs at a constant rate λ. The number of shocks in a time interval of length t is described by a Poisson distribution with the expected number of shocks equal to λ t. Then, the random number of shocks, n, occurring in the interval [0, t] is given by the Poisson distribution: f ( n ) = Pr ( X = n ) =

e − λt ( λ t ) , n = 0,1,2,… > 0, n! n

(3.21)

Since based on this model, the first shock causes component failure, then the component is functioning only when no shocks occur, that is, n = 0. Thus, one can write R ( t λ ) = Pr ( X = 0 ) = e − λt .

(3.22)

Using Equation 3.2, the corresponding pdf can be obtained as f (t ) = λe− λt .

(3.23)

Let us now revisit an interesting property of the exponential distribution in which a failure process represented by the exponential distribution is memoryless. The

102

Reliability and Risk Analysis

TABLE 3.1 Hazard Rate and Applications in Reliability Engineering for Common Probability Distributions

Exponential

Weibull

Gamma

Major Applications in Component Reliability

Hazard Rate h(t)

Distribution

• • • •

h (t ) = λ

β t   α α

h (t ) =

• Used extensively in modeling components • Weakest link model • Corrosion modeling

β −1

When α is continuous: −

• Time between maintenance activities • Time to failure of systems with standby components • Prior distributions in Bayesian analysis

t β

t α −1e x t −   β α  Γ (α ) − x α −1e β dx  0   When α is an integer:

h (t ) =

h (t ) =

Components past the burn-in period Random shocks Electronic components Used in modeling complex systems due to mathematical simplicity

∫

t α −1

β Γ (α ) α

∑

α −1 n=0

((t /β )n / n!)

Normal

Solve from h ( t ) =

f (t ) R (t )

• Life distributions of high stress components • Stress-strength analysis • Tolerance analysis

Lognormal

Solve from h ( t ) =

f (t ) R (t )

• Size distributions of breaks (e.g., in pipes) • Maintenance activities • Prior distributions in Bayesian analysis

Beta

Solve from h ( t ) =

f (t ) R (t )

• Prior distributions in Bayesian analysis

Uniform

 1  for  a ≤ t ≤ b  h (t ) =  b − t  0        otherwise 

Binomial

• Prior distributions in Bayesian analysis • Random number generation • Expert elicitation

k  n   n i  (1 + θ ) −  θ  i=0  k    h ( k ) = 1 +   n k   θ      k

∑

where θ =

−1

• Demand-based failures • Number of failed items in a population

p 1− p

Poisson

 k!  h ( k ) = 1 + k  e µ − 1 − µ   

Geometric

h (k ) =

p 1− p

k

∑ i =1

µ i    i !   

−1

• Homogeneous Poisson Processes • Repairable systems modeling • Rare event modeling • Reliability test design • Maintenance planning

103

Elements of Component Reliability

hazard rate is independent of operating time. Consider the law of conditional probability and assume that an item has survived after operating for a time t. The probability that the item will fail sometime between t and t + ∆t is Pr ( t ≤ T ≤ t + ∆t T > t ) =

e − λt − e − λ(t +∆t ) = 1 − e − λ∆t ,  e − λt

(3.24)

which is independent of t. In other words, the component that has worked without failure up to time t has no memory of its past and remains as good as new. This property can also be easily described by the shock model. That is, at any point along time t, the rate at which fatal shocks occur is the same regardless of whether any shock has occurred up to time t. For cycle-based (discrete failure) components the reliability of the exponential distribution representing the cycle to failure may be expressed as R ( n ) = e −α n = (1 − (1 − e −α ))n .

(3.25)

where n is the cycle of use and α is the failure rate per cycle. Equation (3.25) is the reliability function of the geometric distribution with parameter (1 − e −α ). Note that the failure rate of a geometric distribution is not equal to the failure rate of the corresponding exponential distribution.

3.2.2

Weibull distribution

The Weibull distribution has a wide range of applications in reliability analysis. This distribution covers a variety of shapes. Due to its flexibility for describing hazard rates, all three regions of the bathtub curve can be represented by the Weibull distribution; by using three different Weibull distributions, each can represent one of the bathtub curve regions. It is possible to show that the Weibull distribution is appropriate for a system or complex component composed of a few components or parts whose failure is governed by the most severe defect or vulnerability of its components or parts (i.e., the weakest link model). The functional form of the Weibull hazard rate is a power function: h ( t ) = at b ,  although in a slightly modified form. Using Equation 3.11, the hazard rate, h(t), for the Weibull distribution is: h (t ) =

β t   α α

β −1

, α ,  β > 0, t > 0.

(3.26)

Sometimes the transformation λ = 1/α β is used. In this case Equation 3.26 transforms to h ( t ) = λβ t β −1. This form will be used in Chapter 7. Parameters α and β of the Weibull distribution are the scale and shape parameters, respectively. The shape parameter, β , also relates to the shape of the hazard rate function. • If 0 < β < 1, the Weibull distribution has a decreasing hazard rate that represents the burn-in (early) failure behavior.

104

Reliability and Risk Analysis

• For β = 1, the Weibull distribution reduces to the exponential distribution with the constant failure rate λ = 1/α . • If β > 1, the Weibull distribution represents the wear-out (degradation) region of the bathtub curve having an increasing hazard rate. Different shapes of increasing hazard rate include: • 1 < β < 2: The hazard rate increases less as time increases • β = 2: The hazard rate increases with a linear relationship to time • β > 2: The hazard rate increases more as time increases • β < 3.447798: The distribution is positively skewed (tail to right) • β ≈ 3.447798: The distribution is approximately symmetrical • 3 < β < 4: The distribution approximates a normal distribution • β > 10: The distribution approximates a smallest extreme value distribution. Applications of the Weibull distribution in reliability engineering include: • Corrosion resistance studies. • Time to failure of many types of hardware, including capacitors, relays, electron tubes, germanium transistors, photoconductive cells, ball bearings, and certain motors. • Time to failure of basic elements of a system (components, parts, etc.). Sometimes, a three-parameter Weibull distribution may also be used. This parameterization uses a third parameter, a location parameter, γ . The positive value of this parameter provides a measure of the earliest time at which a failure may be observed. That is, before the time γ the component is failure free. A negative γ  shows that one or more failures could occur prior to the beginning of the reliability data collection. For example, this parameter could count failures occurring during production, in storage, in transit, or any period before the actual use. The hazard rate for this three-parameter Weibull distribution would be represented by h (t ) =

β  t −γ    α α 

β −1

,  β , α > 0, 0 < γ ≤ t < ∞.

(3.27)

Accordingly, the pdf and reliability function become, respectively,

β t −γ  f ( t ) =   α α 

β β −1 −  t − γ      α 

e

,  t ≥ γ , 

(3.28)

and

R (t ) = e

  t −γ −    α

β     

, t ≥ γ .

(3.29)

105

Elements of Component Reliability

There are different forms of discrete (cycle-based) Weibull models discussed in the literature. The form proposed by Padgett and Spurrier (1985) closely corresponds to the continuous Weibull distribution and is expressed as

(

f (n) = 1 − e

−α nβ

)

e

−α

n−1

∑i β i=1

, 

(3.30)

where α is the scale parameter and β is the shape parameter. Accordingly, the hazard rate and reliability functions are represented by, respectively, β

h ( n ) = 1 − e −α n , 

R(n) = e

−α

(3.31)

n−1

∑i β i=1

, with R ( 0 ) = 1.

(3.32)

Note that when β > 0 we have an increasing hazard rate. For β < 0 Equation 3.31 is a decreasing hazard rate, and for β = 0 the distribution reduces to a geometric distribution. Typical values of the cycle-based Weibull parameters in components and devices are α = 0.001 −  0.1,  β = 0.5 −  2. Also note that α is related to the probability of failure at the first cycle. That is, f (1) = 1 − e −α . See Bracquemond and Gaudoin (2003) for other forms of discrete Weibull cycle-based life models. Example 3.3 A printer has a time to failure distribution that can be represented by a Weibull distribution with α = 175  days and β = 4. The warranty for the printer is 50 days. a. Find the probability that the printer survives until the end of the warranty. b. Now assume the printer has survived until the end of the warranty period. Find the conditional reliability of the printer at 70 days after the end of the warranty. c. What region of the bathtub curve is this printer in?

Solution: a. R(t = 50) = 1 − F(t = 50) = 1 − Pr(t ≤ 50). For the Weibull distribution, F ( t α ,  β ) = 1 − e Using α = 175  days and β = 4 , we find

R ( t = 50 ) = e

β  50   −  175 

= 0.993.

β  t −  α

.

106

Reliability and Risk Analysis b. Recall the expression for conditional reliability: R ( x + t t ) =

R (t + x ) . R (t )

Here, t = 50 days, and thus x = 20 days after warranty (so, 70 days total) R ( t > ( 20 + 50 )|t > 50 ) =

Pr ( t > 70 ) 0.975 = = 0.981. Pr ( t > 50 ) 0.993

c. Since β = 4 , the hazard rate is increasing, and thus the printer is experiencing wear-out. Similar to the three-parameter Weibull pdf (Equation 3.28), the smallest extreme value distribution and the largest extreme value distribution are sometimes used as two-parameter distributions. These pdfs take on the following forms. The pdf of the smallest extreme value distribution is given by

1 f (t ) = e δ

t− λ   1 (t − λ ) − e δ δ 

   

,  −∞ < λ < ∞,  δ > 0, − ∞ < t < ∞.

(3.33)

The parameter λ is the location parameter and can take on any value. The parameter δ is the scale parameter and is always positive. The hazard rate for the smallest extreme value distribution is h (t ) =

1 e δ

t− λ δ

, 

(3.34)

which is an increasing function of time, so the smallest extreme value distribution is an increasing hazard rate distribution that can be used as a model for component failures due to aging. In this model, the component’s wear-out period is characterized by an exponentially increasing hazard rate. Clearly, negative values of t are not meaningful when representing time to failure. The Weibull distribution and the smallest extreme value distribution are closely related to each other. If an r.v. X follows the Weibull distribution as given in Table 2.10, the r.v. T using the transformation t = ln ( x ) follows the smallest extreme value distribution with parameters

λ = ln (α ) , δ =

1 . β

(3.35)

The two-parameter largest extreme value pdf is given by

1 f (t ) = e δ

t− λ   − 1 (t − λ ) − e δ  δ 

   

,  −∞ < λ < ∞,  δ > 0, − ∞ < t < ∞.

(3.36)

The largest extreme value distribution, although not very useful for component failure behavior modeling, is useful for estimating natural extreme phenomena. For

107

Elements of Component Reliability

further discussions regarding extreme value distributions, see Johnson et al. (1995), Castillo (1988), and Gumble (1958).

Example 3.4 The maximum demand for electric power during a year is directly related to extreme weather conditions. An electric utility has determined that the distribution of maximum power demands, t, can be modeled by the largest extreme value distribution with λ = 1,200 MW and δ = 480 MW . Determine the unreliability of the installed power, represented by the probability (per year) that the demand will exceed the utility’s maximum installed power of 3,000 (MW).

Solution: Since this is the largest extreme value distribution, we should integrate Equation 3.36 from 3,000 to +∞: ∞

Pr ( t > 3,000 ) =

∫ f ( t ) dt = 1 − e

  t−λ     − δ    − e  

.

3,000

Since t − λ 3,000 − 1,200 = = 3.75, δ 480 then Pr(t > 3,000) = 0.023.

3.2.3

gaMMa distribution

The gamma distribution was introduced in Chapter 2 as a generalization of the exponential distribution representing the sum of α independent exponential variables. Recalling the simple shock model considered in Section 3.2.1, one can expand this model for the case when a component fails after being subjected to α successive random shocks arriving at a constant rate (assuming integer α ). In this case, the time to failure distribution of the component follows the gamma distribution. Examples of its application include the distribution of times between recalibration of an instrument that needs recalibration after α uses, time between maintenance of items that require maintenance after α uses, and time to failure of a system with standby components, having the same exponential time to failure distribution with β = 1/λ where λ is the exponential parameter. It can be seen that β is the mean time to occurrence of a single event. The gamma distribution has two parameters, α (shape parameter) and β (scale parameter). The gamma cdf and reliability function, in the general case, do not have closed forms.

108

Reliability and Risk Analysis

When the shape parameter α is an integer, the gamma distribution is known as the Erlangian distribution. Here, the reliability and hazard rate functions can be expressed in terms of the Poisson distribution as n

R (t ) =

α −1

∑ n=0

t

 t  −β  β  e ,  n!

(3.37)

t α −1

h (t ) =

n

β α Γ (α )

∑

t   α −1   β  n=0 n!

.

(3.38)

In this case, α shows the number of shocks required before a failure occurs and β represents the mean time to occurrence of a shock. The gamma distribution represents a decreasing hazard rate for α < 1, a constant hazard rate for α = 1, and an increasing failure rate for α > 1. Thus, the gamma distribution can represent each of the three regions of the bathtub curve. Example 3.5 The mean time to adjustment of an engine in an aircraft is E(T) = 100 flight hours (assume time to adjustment follows the exponential distribution). Suppose there is a maintenance requirement to replace certain parts of the engine after three consecutive adjustments. a. What is the distribution of the time to replace? b. What is the probability that a randomly selected engine does not require part replacement for at least 200 flight hours? c. What is the mean time to replace?

Solution: a. Use gamma distribution for T with α = 3,  β = 100. n

b. For the gamma distribution, R ( t ) =

2

∑ n=0

 t 

 t  −  100    e 100  ,  n! 0

 200  −2   e ( 2)1 e −2 ( 2)2 e −2 100  = + + , 0! 1! 2! = 0.135 + 0.271 + 0.271 = 0.677. c. Mean time to replace is E(T) = αβ = 3(100) = 300 flight hours.

Elements of Component Reliability

3.2.4

109

norMal distribution

The normal distribution is a basic distribution of statistics. The popularity of this distribution in reliability engineering can be explained by the central limit theorem. According to this theorem, the sum of a large number, n, of independent r.v.’s approaches the normal distribution as n approaches infinity. For example, consider a sample of many randomly generated observations such that the observations are independent of each other. If we generate multiple samples of this type, the central limit theorem says that as the number of these sample increase, the pdf representing the arithmetic mean of the samples will approach a normal distribution. The normal distribution is an appropriate model for many practical engineering situations. Since a normally distributed r.v. can take on a value from the (−∞, +∞) range, it has limited applications in reliability problems that involve time to failure estimations because time cannot take on negative values. However, for cases where the mean μ is positive and is larger than σ by several folds, the probability that the r.v. T takes negative values can be negligible. For cases where the probability that T takes negative values is not negligible, the truncated normal distribution can be used (see O’Connor et al., 2019). The normal distribution hazard rate is always a monotonically increasing function of time, t, so the normal distribution is an increasing hazard rate distribution. Thus, the normal distribution can be used to represent the high stress wear-out region of the bathtub curve. The normal distribution is also a widely used model representing stress and/or strength in the framework of the stress-strength reliability models, which are time-independent reliability models. The normal distribution can also model simple repair or inspection tasks if they have a typical duration and variance symmetrical about the mean.

3.2.5

lognorMal distribution

The lognormal distribution is widely used in reliability engineering. The lognormal distribution represents the distribution of an r.v. whose logarithm follows the normal distribution. The lognormal distribution is commonly used to represent the occurrence of certain events in time or space whose values span by several folds or more than one order of magnitude. For example, an r.v. representing the length of time required for a repair often follows a lognormal distribution, because depending on the skills of the repair technician the time to finish the job might be significantly different. This model is suitable for failure processes that result from many small multiplicative errors. Specific applications of this distribution include time to failure of components due to fatigue cracks. Other applications of the lognormal distribution are associated with failures attributed to maintenance activities and distribution of cracks initiated and grown by mechanical fatigue. The distribution is also used as a model representing the distribution of particle sizes observed in breakage processes and the life distribution of some electronic components. Data that follows the lognormal distribution can typically be identified quickly as some data points may be orders of magnitudes apart. In Bayesian reliability analysis, the lognormal distribution is a popular model to represent the prior distributions. We discuss this topic further in Chapter 5.

110

Reliability and Risk Analysis

The hazard rate for the lognormal distribution initially increases over time and then decreases. The rate of increase and decrease depends on the values of the parameters µ and σ . In general, this distribution is appropriate for representing time to failure for a component whose early failures (or processes resulting in failures) dominate its overall failure behavior. Example 3.6 The time to failure (in hours) of an experimental laser device is given by the lognormal distribution, with parameters µ = 3.5 hours and σ = 0.9 . Find the MTTF for the laser device. Then, find the reliability of the laser device at 25 hours.

Solution: The MTTF is found as the mean of the lognormal distribution with the given parameters:

MTTF = E ( t ) = e

 σ2  µ−  2  

=e

 0.92   3.5−  2  

= 49.65 hours.

Solving for reliability,  ln x − µ  R ( t = 25) = 1 − F ( t = 25) = 1 − Φ  ,  σ 

 ln 25 − 3.5  Φ = Φ ( −0.312 ) ,   0.9 From the lookup table in Appendix A, Φ −1 ( −0.312 ) = 0.377 Thus, R ( 25) = 1 − 0.377 = 0.623.

3.2.6

beta distribution

The beta distribution is often used to model parameters that are constrained in an interval. The distribution of a probability parameter 0 ≤ p ≤1 is popular with the beta distribution. This distribution is frequently used to model proportions of a subclass of events in a set. An example of this is the likelihood ratios for estimating uncertainty. The beta distribution is often used as a conjugate prior in Bayesian analysis for the Bernoulli, binomial and geometric distributions to produce closed form posteriors. The hazard rate for the Beta distribution may be expressed by h (t ) =

t α −1 (1 − t ) , B (α ,  β ) − B ( t α ,  β )

(3.39)

111

Elements of Component Reliability

where B (α ,  β ) is the beta function, B ( t α ,  β ) is the incomplete beta function. The reliability function of the beta distribution is often used in reliability testing which shows the probability that a reliability target is met. It is also used in Bayesian analysis as the prior pdf of reliability during pass-fail reliability demonstration tests. Example 3.7 The probability of failure on demand, pFOD , for a backup diesel generator is modeled by the beta distribution with α = 3 and β = 150. a. Find the mean and the variance of the probability of failure on demand. b. Use the mean value and find the probability that the generator fails exactly two times in the next 50 times it is demanded.

Solution: a. Using the properties of the beta distribution from Chapter 2, E ( pFOD ) = Var ( pFOD ) =

α 3 = = 0.0196, α + β 3 + 150 αβ

(α + β ) (α + β + 1) 2

=

3 ⋅150 = 1.25 × 10 −4 . (153)2 (154 )

b. The mean value from part (a) gives a probability of failure on demand. This becomes the parameter, p, of a binomial distribution. We are told we have n = 50 demands, and want to know the probability that x = 2 failures. Using the binomial pdf:  n f ( x n ,  p ) =   p x (1− p)n − x = f ( x n ,  p ) ,  x  50  Pr ( x = 2 ) =   0.01962 (1 − 0.0196)50 − 2 = 1225⋅ 0.01962 (0.9804) 48 = 0.182.  2 

3.3

EXERCISES

3.1 Assume that T, the random variable that denotes life in hours of a specified component, has a cumulative density function (cdf) of  100 , t ≥ 100,  1− F ( t ) =   t   0, t < 100.  Determine the following: a. Pdf f ( t ) b. Reliability function R ( t ) c. MTTF

112



Reliability and Risk Analysis

113

Elements of Component Reliability

3.10 The reliability of a propellor can be represented by the following expression: 2

 t R ( t ) =  1 −  , 0 ≤ t ≤ t0  t0 





where t0 is the maximum life of the propellor a. What region of the bathtub curve is this propellor in? b. Compute MTTF if t0 = 3,000 hours . 3.11 The following is the time to failure pdf of a new appliance, f ( t ) = 0.1(1 + 0.05t ) , t ≥ 0. 3



3.12 3.13



3.14

3.15

3.16

3.17

Determine the fraction of appliances that will fail up to the warranty period of 18 months. Suppose X has the exponential pdf f ( x ) = λ e − λ x for x > 0, and f ( x ) = 0 for x ≤ 0. Find Pr(x > ( a + b ) x > a),  a,  b > 0. A manufacturer uses the exponential distribution to model the number of cycles to for a product. The product has λ = 0.003 failures/cycle, a. What is the mean cycle to failure for this product? b. If the product survives for 300 cycles, what is the probability that it will fail sometimes after 500 cycles? If operational data show that 1,000 components have survived 300 cycles, how many of these would be expected to fail after 500 cycles? Life of an aging device can be described by a Weibull distribution with the shape parameter of 2.15 and a MTTF of 25,500 hours. Assuming that a failed device is replaced with a new one that does not fail, determine how many devices among the 30 operating ones are expected to fail in 1,500 hours. Time to failure of a relay follows a Weibull distribution with α = 10 years, β = 0.5. Find the following: a. Pr(failure after 1 year) b. Pr(failure after 10 years) c. MTTF A medical device has a time to failure distribution given by a Weibull distribution with a shape parameter of 1.4 and E(t) = 50,000 hours. a. Determine the reliability at 1 year. b. Calculate the conditional reliability of the device at 15,000 hours if the device has already survived 5,000 hours. The MTTF of a certain type of small motor is 10 years, with a standard deviation of 2 years. The manufacturer replaces free of charge all motors that fail while under warranty. If the manufacturer is willing to replace only 3% of the motors that fail, what warranty period should they offer? Assume that the time to failure of the motors follows a normal distribution.

114

Reliability and Risk Analysis



REFERENCES Bracquemond, C., and O. Gaudoin, “A Survey on Discrete Lifetime Distributions.” International Journal of Reliability, Quality and Safety Engineering, 10(1), 69–98, 2003. Castillo, E., Extreme Value Theory in Engineering. Academy Press, San Diego, CA, 1988. Gumble, E. J., Statistics of Extremes. Columbia University Press, New York, 1958. Gupta, P. L., R. C. Gupta, and R. C. Tripathi, “On the Monotonic Properties of Discrete Failure Rates.” Journal of Statistical Planning and Inference, 65(2), 255–268, 1997. Johnson, N. L., S. T. Kotz, and N. Balakrishnan, Continuous Univariate Distributions. John Wiley & Sons, Vol. 2, New York, 1995. O’Connor, A.N., M. Modarres, and A. Mosleh, Probability Distributions Used in Reliability Engineering, DML International, Washington, DC, 2019. Padgett, W. J., and J. D. Spurrier, “On Discrete Failure Models.” IEEE Transactions on Reliability, 34(3), 253–256, 1985.

4

Basic Reliability Mathematics Statistics

4.1 INTRODUCTION Statistics is the process of collecting, analyzing, organizing, and interpreting data. Data are used to support decisions about a population of interest. Data can be collected about a full population, but more often, the collected data are a sample from the population of interest. A set of observations from a distribution of an r.v. is called a sample. The number of observations in a sample is called the sample size. In the framework of classical statistics, a sample is usually composed of exchangeable, independently, and identically distributed (i.i.d.) observations. From a practical point of view, this assumption means that elements of a sample are obtained individually and without regards to other samples and under the same conditions. Now, let’s discuss two aspects of statistics of primary relevance. The first, descriptive statistics, is concerned with descriptive properties of observed data. The second, inferential statistics (or statistical inference) is concerned with using data to make conclusions about a population. In reliability, for example, we use descriptive statistics such as the mean and variance of observed data to characterize failure data, and inferential statistics to establish a probability distribution representing time to failure of a population of identical components. Caution: statistical terminology differs from similar terminology used in machine learning. In machine learning, the term inference is used to refer to making a prediction using a trained model; in this context, determining properties of the model is called learning (rather than inference), and using the model for prediction is called inference (rather than prediction). In statistical inference, we can ask deductive questions, such as “given a population, what will a sample look like?” Alternatively, we can ask inductive questions, such as “given a sample, what can be inferred about a population?” Since a sample is a subset of the population, it is necessary to make corrections based on sample characteristics, randomness, assumptions, and data collecting characteristics such as censoring. We will describe several corrections later in this chapter. A sample of data provides the basis for statistical inference about the underlying distribution of the parent population. Typically, the data come from special tests, experiments, operational data, or practical uses of a limited set of items. Generally, it is assumed that each sample point is i.i.d., but it is prudent to verify this before applying methods blindly. Each observed value is considered a realization (or observation) of some hypothetical r.v.: that is, a value that the r.v. can take on. DOI: 10.1201/9781003307495-4

115

116

Reliability and Risk Analysis

4.2 DESCRIPTIVE STATISTICS Two important measurements related to observed samples are mean and variance. These are closely related to the concepts described for probability distributions in Section 2.6.2. One can define the sample arithmetic mean and sample variance as the respective expected values of a sample of size n from the distribution of X, namely, x1 , …, x n as follows: 1 E(X ) = x = n

n

∑x , 

(4.1)

i

i=1

and Var ( X ) = s 2 =

1 n

n

∑(x − x ) .  2

i

(4.2)

i=1

When sample data vary by orders of magnitude, it is customary to use geometric mean. In this case, the logarithm of the sample data is used to represent xi in Equations 4.1 and 4.2. Equation 4.2 can be used to estimate the variance if the data represent the complete population. However, if the data come from a sample, the estimator of variance (Equation 4.2) is biased, since x is estimated from the same sample. It can be shown that this bias can be removed by multiplying Equation 4.2 by n/(n−1):

Var ( X ) = s 2 =

1 n −1

n

∑(x − x ) . i

2

(4.3)

i=1

The concept of median also applies to a sample, where the sample median is simply the midpoint of a set of data. Additional topics fall within the domain of descriptive statistics, including moments, sample covariance, and more. The reader is referred to statistics textbooks for full coverage. Example 4.1 A sample of eight manufactured shafts is taken from a plant lot. The diameters of the shafts are 1.01, 1.08, 1.05, 1.01, 1.00, 1.02, 0.99, and 1.02 inches. Find the sample mean and variance.

Solution: E(X ) = x =

1 n

n

∑x = 1.0225. i

i=1

117

Basic Reliability Mathematics Since these data are samples, we use the unbiased form the variance. Var ( X ) =

1 n −1

n

∑(x − x ) i

2

= 0.0085.

i=1

4.3 EMPIRICAL DISTRIBUTIONS AND HISTOGRAMS When studying distributions, it is convenient to start with some preliminary procedures useful for data visualization, editing, and detecting outliers by constructing empirical distributions and histograms. Such preliminary data analysis procedures are useful to support understanding data, to illustrate the data (sometimes the data speak for themselves), to identify suitable distributions to consider, and to support other types of analysis (goodness of fit testing, for instance). To illustrate some of these procedures, consider the data set composed of observed times to failure of 100 identical electronic devices, given in the first two columns of Table 4.1. Notice that the data are already aggregated into bins. The measure of interest here is the relative frequency associated with each interval of time to failure. This can be obtained using Equation 2.10, that is, by dividing each interval frequency by the total number of devices tested. Applying Equation 2.10 produces the empirical probability distribution in the third column of Table 4.1. TABLE 4.1 Example of Binned Data with Corresponding Empirical Probability Distribution Time Interval, T (hour) 0–100 100–200 200–300 300–400 400–500 500–600 600–700 700–800 800–900 900–1,000 Total:

Observed Frequency

Relative Frequency, ne n

35 26 11 12 6 3 4 2 0 1 100

0.35 0.26 0.11 0.12 0.06 0.03 0.04 0.02 0.00 0.01 1.0

Example 4.2 Consider the observed time to failure data for an electronic device below. It is believed that the data come from a time to failure process, T, that can be represented by an exponential distribution with parameter λ = 0.005 hour −1 . Determine the expected frequency of failures in each time interval.

118

Reliability and Risk Analysis

Solution: The probability (relative frequency) that T takes values between 0 and 100 hours is 100

Pr(0 < T < 100) =

∫ 0.005e

−0.005t

dt,  

0

Pr(0 < T < 100) = [1 − e −0.005t ]100 0 = 0.393. Summing the observed frequency column shows that 100 data points were collected. By multiplying the probability by the total number of devices observed (100), we can determine the expected frequency for the interval 0–100. The expected frequency for the 0–100 interval is  0.393 ⋅ 100 = 39.3. The results for the rest of the intervals are shown below. Interval, T (hour)

Observed Frequency

Exponential Expected Frequency (Count)

Exponential Expected Relative Frequency

0–100

35

39.3

0.393

100–200 200–300 300–400 400–500 500–600 600–700 700–800 800–900 900–1,000

26 11 12 6 3 4 2 0 1

23.8 14.5 8.8 5.3 3.2 2.0 1.2 0.7 0.4

0.238 0.145 0.088 0.053 0.032 0.020 0.012 0.007 0.004

A comparison of the observed and expected frequencies of each interval reveals differences as large as 4.3 failures.

Figure 4.1 illustrates the respective histogram of the data from Example 4.2 and its comparison to the exponential distribution with λ = 0.005 hour −1.

t

FIGURE 4.1 Observed frequencies (histogram) and expected frequencies from Example 4.2.

119

Basic Reliability Mathematics

4.4

PARAMETER ESTIMATION: POINT ESTIMATION

Statistical inference involves using data to identify the distribution of an r.v. We can use a sample of failure times of an item, t1 , …, tn , to estimate, for example, λ , the parameter of an exponential distribution representing time to failure of the item. In this case, we are inferring a general distribution from a specific sample of data. Since a sample is one possible realization of data from the population, estimating the model parameter inherently involves uncertainty. There are many methods for conducting this parameter estimation, some in the domain of frequentist statistics and others in the domain of Bayesian statistics. Let f ( x θ ) denote the pdf of r.v. X where θ represents an unknown parameter or a vector of parameters. Let x1 , …, x n denote a sample from f ( x θ ) . In frequentist statistics, the parameter θ is considered to be fixed but unknown, and a random sample of data is used to estimate it. In Bayesian statistics, the parameter θ is considered to be an r.v., and a fixed set of data (or evidence) is used to update our prior knowledge of the parameter. The contrasting viewpoints on data and parameters between the frequentist and Bayesian are critical in interpreting and formulating the likelihood function in the maximum likelihood estimation and Bayesian estimation procedures that will be discussed later in Sections 4.4.3 and 4.4.4. Point estimation and interval estimation are the two basic kinds of estimation procedures. Point estimation uses a data set to obtain a single number that represents the most likely value of a parameter of the distribution function or other characteristic of the underlying distribution of interest. Point estimation does not provide information about the uncertainty in that number. Uncertainty is expressed using interval estimation, with either confidence intervals (frequentist) or credible intervals (Bayesian) which will be discussed later in this chapter. Now, let’s discuss terminology. Suppose we are interested in estimating a singleparameter distribution f ( x θ ) based on a random sample x1 , …, x n . Let g ( x1 , …, x n ) be a single-valued (simple) function of x1 , …, x n . It follows that g ( x1 , …, x n ) is also an r.v., which is called a statistic. A point estimate of that function is obtained by using an appropriate statistic and calculating its value based on the sample data. The statistic (as a function) is called the estimator and its numerical value is called the estimate. Consider the basic properties of point estimators. An estimator g ( x1 , …, x n ) is said to be an unbiased estimator for θ if its expectation coincides with the value of the parameter of interest θ. That is, E  g ( x1 , …, x n ) = θ for any value of θ. Thus, the bias in the estimator is the difference between the expected value of an estimator and the true parameter value itself. It is obvious that the smaller the bias, the better the estimator is. Another desirable property of an estimator g ( x1 , …, x n ) is the property of consistency. An estimator g ( x1 , …, x n ) is said to be consistent if, for every ε > 0,

(

)

lim Pr g ( x1 , …,  x n ) − θ < ε = 1.

n→∞ 

(4.4)

This property implies that as the sample size n increases, the estimator g ( x1 , …, x n ) gets closer to the true value of θ. In some situations, several unbiased estimators

120

Reliability and Risk Analysis

can be found. A possible procedure for selecting the best one among the unbiased estimators can be based on choosing the one having the least variance. An unbiased estimator t of θ, having minimum variance among all unbiased estimators of θ, is called an efficient estimator. Another estimation property is sufficiency. An estimator g ( x1 , …, x n ) is said to be a sufficient statistic for the parameter if it contains all the information in the sample x1 , …, x n . In other words, the sample x1 , …, x n can be replaced by g ( x1 , …, x n ) without loss of any information about the parameter of interest. Several methods of estimation are considered in statistics. In the following section, some of the most common methods are briefly discussed. Estimated parameters will be denoted with a hat, i.e., θˆ denotes the estimate for parameter θ .

4.4.1

Method of MoMents

The method of moments is an estimation procedure based on empirically estimated sample moments of the r.v. According to this procedure, the sample moments are equated to the corresponding distribution moments. That is, x and s2 obtained from the equations in Section 4.2 can be used as the point estimates of the distribution mean, µ, and variance, σ 2. The solutions of these equalities provide the estimates of the distribution parameters.

4.4.2

linear regression

Another widely used method for parameter estimation is a nonstatistical optimization approach by linear regression, often used in conjunction with probability plotting methods. Because regression has many uses in addition to parameter estimation, we cover it separately in Section 4.7. We will return to the use of linear regression with the probability distribution plotting methods in Chapter 5.

4.4.3 MaxiMuM likelihood estiMation This method is one of the most widely used methods of estimation. Consider a continuous r.v. X with pdf f ( X θ ). Also consider a random set of data (evidence) Ei which provides information about a sample of values of X. The probability of occurrence of the sample evidence Ei depends on the parameter θ . The likelihood of this parameter is given by a likelihood function, L (θ ) that is proportional to the probability of the observed evidence,  L (θ ) ∝ Pr  



∩ E θ  . i

(4.5)

i

Let our evidence Ei be a random sample x1 , …, x n of size n taken from the distribution of X. Assuming these observations are statistically independent, the likelihood of obtaining this particular set of sample values is proportional to the joint occurrence of n independent random events x1 , …,  x n . That is, the likelihood function is

121

Basic Reliability Mathematics

proportional to the probability (for a discrete r.v.) or the relative frequency of the pdf (for a continuous r.v.) of the joint occurrence of x1 , …, x n : L (θ x1 , …, x n ) ∝

n

∏ f ( x θ ) = f ( x θ )… f ( x θ ). 1

i

n

(4.6)

i=1

Note that the pdf may be viewed as a function of r.v. X with the fixed but unknown pa rameter θ , i.e., f ( x θ ). In this case the likelihood function is written as L (θ x1 ,  …,  x n ) . However, the pdf can alternatively be viewed as a function of variable θ with fixed x, i.e., f (θ x ), in which case the likelihood function is written as L ( x1 ,  …,  x n θ ). The former view is how the MLE approach expresses the likelihood function, whereas the latter represents the Bayesian parameter estimation method’s view of the likelihood function. The maximum likelihood estimate (MLE), is the parameter value that maximizes the likelihood function, L (θ x1 , …, x n ).

θˆ = arg max [ L (θ )] ,  subject to θ  contraints.

(4.7)

The standard way to find a maximum of a parameter in MLE is to optimize the likelihood function. The simplest way to do this is to calculate the first partial derivative of L (θ ) with respect to each parameter and equate it to zero. This yields the equation(s) ∂ L (θ x1 , …, x n ) = 0, ∂θ

(4.8)

from which the MLE θˆ can be obtained. Due to the multiplicative form of Equation 4.6, in many cases it is more convenient to maximize the logarithm of the likelihood function, the log-likelihood Λ (θ x ): ∂log L (θ x1 , …,  x n ) ∂Λ (θ ) = = 0. ∂θ ∂θ

(4.9)

Since the logarithm is a monotonic transformation, the estimate of θ obtained from this equation is the same as that obtained from Equation 4.8. Sometimes, Equations 4.8 or 4.9 can be solved analytically, but often must be solved numerically. The optimum values of the unknown (but fixed) parameters, θˆ, are called point estimates or best estimates of the parameters in the classical (frequentist) treatments of probability. For example, it can be shown that for complete data of n times to failure, ti , believed to come from a normal distribution, the MLE parameters of the normal distribution are: 1 µˆ = n

n

∑t ,  i

i=1

(4.10)

122

Reliability and Risk Analysis

1 σˆ 2   = n

n

∑ (t − µ ) . 2

(4.11)

i

i=1

Usually, the unbiased estimator form of Equation 4.11 is used, as explained in Section 4.2: 1 σˆ 2   = n −1

n

∑ (t − µ ) . 2

(4.12)

i

i=1

To express the confidence over the estimation of the unknown fixed parameters, we should also express a confidence interval associated with a confidence level (likelihood) within which the true value of the unknown parameter resides. This concept will be discussed in Section 4.5. Example 4.3 Consider a sample t1 ,  …,  tn of n times to failure of a component whose time to failure is known to be exponentially distributed. Find the MLE of the failure rate, λ.

Solution: Using Equations 4.6 and 4.9, one can obtain n

L = L (λ | t1 ,   …,  tn ) =

∏λ  e

− λ ti

=λ e n

−λ

n

∑ti i=1

,

i=1

Λ = ln L = n ( ln λ ) − λ

n

∑t , i

i=1

∂Λ n  =  − λ ∂λ

λˆ =

n

∑t

= 0,

i 

i=1

n

∑

n

. ti

i=1

 ∂2 Λ  n Recalling the second-order condition  2  < 0, since − 2 < 0 , the estimate λ  ∂λ  λˆ

λˆ is indeed the MLE for the problem considered.

123

Basic Reliability Mathematics

4.4.4

bayesian ParaMeter estiMation

Bayesian parameter estimation uses the subjective interpretation of probability and Bayes’ Theorem to update the prior state of knowledge about the unknown parameter of interest, which is treated as an r.v. Recall Bayes’ Theorem, which was introduced in Chapter 2 in the context of discrete variables as Equation 2.35. The continuous form of Bayes Theorem is: f1 (θ E ) =

f0 (θ ) L ( E θ )

∫

f0 (θ ) L ( E θ )dθ

,

(4.13)

where f0 (θ ) is the prior pdf of r.v. θ , f1 (θ E ) is the posterior pdf of θ , and L ( E θ ) is the likelihood function from Equation 4.6. The denominator of Equation 4.13, once integrated, is a constant. Now, consider new evidence of the form of a random sample x1 , …, x n of size n taken from the distribution of X. We want to use this evidence for estimating parameter(s) of a distribution. In this instance, Bayes’ Theorem can be expressed by f1 (θ x1 , …, x n ) =

f0 (θ ) L ( x1 , …,  x n θ )

∫ f (θ ) L(x , …,  x θ )dθ 0

Pr1 (θ x1 , …,  x n ) =

1

, 

n

Pr0 (θ ) Pr( x1 , …,  x n θ )

∑

n

(4.14)

Pr0 (θ i ) Pr( x1 , …,  x n θ i )

, 

(4.15)

i=1

respectively for continuous or discrete distributions, where f0 (θ ) and Pr0 (θ ) are the prior pdfs, and L ( x1 , …, x n θ ) and Pr ( x1 , …,  x n θ ) are the likelihood functions representing the observed data and information (evidence). The prior distribution is the probability distribution of the r.v. θ , which captures our state of knowledge of θ prior to the data x1 , …, x n being observed. It is common for this distribution to represent soft evidence or probability intervals about the possible range of prior values of θ . If the distribution is dispersed it represents a case where little is known about the parameter θ . If the distribution is concentrated (has higher density) in an area, then it reflects a good knowledge about the likely values of θ . This Bayesian estimation process is illustrated in Figure 4.2. The relationship with MLE also becomes clear by observing Figure 4.2. In Bayesian analysis we combine the prior distribution of the unknown random parameter(s) with the likelihood function representing the observed data (evidence) by considering the data as fixed evidence. We use it to obtain f1 (θ x1 , …,  x n ) or Pr1 (θ x1 , …,  x n ) , the posterior pdf or probability distribution of the parameter(s). To obtain a single statistic akin to the MLE, we use either the mean or median of the distribution as the Bayesian estimate.

124

Reliability and Risk Analysis

Model for Data

Likelihood L(Data| )

MLE

Bayesian Inference

Posterior ( | Data)

Data Prior

FIGURE 4.2

Bayesian inference method.

Note that we used L (θ x1 , …, x n ) in MLE but use L ( x1 , …,  x n θ ) in Bayesian estimation. As noted before, this is because the MLE approach is intended to obtain an estimation of the unknown but fixed parameters (regardless of any prior knowledge), using only observed random samples of data (evidence). In essence, because the frequentist approach does not consider the prior information, the MLE approach treats L (θ E ) and L ( E θ ) as equal. To express our uncertainties about the estimated parameters, Bayesian estimation develops the credible interval which shows the probability that the r.v. θ falls within an interval of values. This topic is further discussed next. Bayesian estimation will be examined in far more detail in Chapter 5 in the context of reliability parameter estimation.

4.5 PARAMETER ESTIMATION: INTERVAL ESTIMATION Interval estimation is used to quantify the uncertainty about the parameter θ due to sampling error (e.g., a limited number of samples). Interval estimates are expressed as either confidence intervals (frequentist) or credible intervals (Bayesian). Readers should be careful in understanding the differences between the meaning and use of these intervals in decision making. Typical intervals of interest are 99%, 95%, and 90%. Note that interval estimation does not address uncertainty due to inappropriate model selection or invalid assumptions. Care must be taken to avoid these other issues by verifying assumptions and model correctness.

4.5.1

confidence intervals

The confidence interval is the frequentist approach to express uncertainties about the estimated parameters. The main purpose is for finding an interval with a high probability of containing the true but unknown value of the parameter θ . For the parameter θ describing the distribution of an r.v. X, a confidence interval is estimated based on a sample x1 , …, x n  of size n from the distribution of X. Consider two r.v.s θ l ( x1 , …,  x n ) and θ u ( x1 , …,  x n ) chosen in such a way that the probability that interval [θ l , θ u ] contains the true parameter θ is Pr[θ l ( x1 , …,  x n ) < θ < θ u ( x1 , …,  x n )] = 1 − γ .

(4.16)

Basic Reliability Mathematics

125

The interval [θ l , θ u ] is called a k% confidence interval (specifically, a two-sided confidence interval) for the parameter θ. The term k% = 100 (1 – γ ) % is the confidence coefficient or confidence level, and the endpoints θ l , and θ u  are called the k% lower and upper confidence limits of θ , respectively. In the frequentist approach, these interval endpoints are estimated from sample data. The confidence interval is interpreted as follows. If many confidence intervals are generated this way (i.e., from many sample sets from the same population), then k% of those intervals will contain the true value of the fixed parameter θ . Reversely, this also means that 100γ % of the generated intervals will not contain the true value of θ . Therefore, we have a k% confidence that any generated interval contains the true value of θ . It is critical to note that the expressed confidence level k% is about the interval not the unknown parameter. See Figure 4.3 for an illustration of this concept. Sometimes, frequentist intervals are sometimes misinterpreted to mean that there is a k% probability that the parameter will be in that interval. This probability is not given by the confidence interval (rather, it is given by the Bayesian credible interval). In the frequentist approach, it is not appropriate to make statements about the probability of a parameter; only the Bayesian approach does this. Each row in Figure 4.3 is a sample x1 , …, x n  of n points from the distribution with unknown parameter θ (in this case, mean µ). The square indicates the sample mean and the whiskers represent the 70% confidence interval. The white squares indicate intervals that contain the true population mean, the black squares indicate intervals that do not. Because these are frequentist confidence intervals, 70% of the generated intervals contain the true population mean. It is also possible to construct one-sided confidence intervals. The case when θ > θ l with the probability of 1 – γ , θ l is called the one-sided lower confidence limit

FIGURE 4.3

Illustration of how confidence intervals are generated from a set of samples.

126

Reliability and Risk Analysis

for θ . The case when θ < θ u  with probability of 1 – γ , θ u is called the one-sided upper confidence limit for θ . Consider a typical example illustrating the basic idea of confidence limits for the parameters of a normal distribution. First, consider a procedure for constructing confidence intervals for µ when σ is known. Let x1 , …, x n be a random sample from the normal distribution, X ~ N ( µ , σ ) . It can be shown that the sample mean x has the σ  ( x – µ ) has the  normal distribution X ~ N  µ ,  . By using the Z transformation,   n σ/ n standard normal distribution. Using this distribution, one can write   x−µ   Pr  z γ  ≤ σ ≤ z γ   = 1 − γ ,  1−  2   2    n

(4.17)

where z(γ /2) is the 100 (γ 2 ) % value of the standard normal distribution, which can be obtained from Table A.1. After simple algebraic transformations, and recalling that the normal distribution is symmetric and thus z(γ /2) = −z(1−γ / 2) , the inequalities inside the parentheses of Equation 4.17 can be rewritten as  σ σ  Pr  x − z γ  ≤ µ ≤ x + z γ  = 1− γ . 1− 1− n n    2   2 

(4.18)

Equation 4.18 provides the symmetric (1 – γ) confidence interval on parameter µ . Generally, a two-sided confidence interval is wider for a higher confidence level (1 – γ). As the sample size n increases, the confidence interval becomes narrower for the same confidence level (1 – γ). When σ is unknown and is estimated from the sample as σˆ = s, or in cases when ( x – µ ) follows the the sample is small (e.g., n < 30), it can be shown that the r.v. s/ n Student’s t-distribution with ν = [ n – 1] degrees of freedom. Transforming this as above, and recalling that the t-distribution is symmetric, the confidence interval on µ is given by:  s s  Pr  x − t γ  < µ < x + t γ  = 1 − γ ,  1−  1−  n n    2  2

(4.19)

γ is the 100  1−  th percentile of the one-tailed t-distribution with ν = [ n – 1]  2 degrees of freedom (see Table A.2 or use the Excel command t.inv(1 − γ /2, n–1)). Similarly, confidence intervals for σ 2 for a normal distribution for a sample with variance s 2 can be obtained as

where t

γ  1−  2

127

Basic Reliability Mathematics

( n − 1) s 2 < σ 2 < ( n − 1) s 2 ,  2 2

χ

γ 1−   2 

[ n − 1]

χ  γ  [ n − 1]

(4.20)

  2

γ th percentile of a χ2 distribution with df = [ n – 1] 2 degrees of freedom. They can be found using Table A.3, the Excel command chisq. inv(1 – γ , df), or in MATLAB using chi2inv(1 – γ , df). Confidence intervals for other distributions will be further discussed in Chapter 5.

where χ 2

γ  1−  2

4.5.2

[ n − 1] is the 100  1−

credible intervals

The Bayesian analog of the classical confidence interval is known as the credible interval or Bayesian probability interval. To construct a Bayesian probability interval, the following relationship based on the posterior distribution is used: Pr (θ l < θ ≤ θ u ) = 1 − γ .

(4.21)

The interval [θ l , θ u ] is the k% = 100 (1 – γ ) % credible interval (specifically a twosided credible interval) for the parameter θ, and θ l and θ u are the k% lower and upper credible limits of θ . Interpretation of the credible interval is simple and intuitive. The Bayesian credible interval directly gives the probability that an unknown parameter θ falls into the interval [θ l , θ u ]. At k% probability level, the credible interval gives a range which has a k% chance of containing θ , as shown in Figure 4.4. Like the confidence interval, the Bayesian credible interval can be calculated as a two-sided or onesided interval. Note that the for the confidence interval, the k% confidence level is a property of the confidence interval estimation (not a property of θ the unknown parameter of interest), whereas the k% credible interval has direct relevance to the parameter θ . The calculation of the credible interval follows directly from the posterior distribution. We will illustrate this in Example 4.4

FIGURE 4.4 Illustration of Bayesian credible intervals for a parameter θ .

128

Reliability and Risk Analysis

Example 4.4 A capacitor has a time to failure that can be represented by the exponential distribution with a failure rate λ . The mean value and uncertainty about λ   are modeled through a Bayesian inference with the posterior pdf for λ   expressed as a normal distribution with a mean of 1 × 10 −3 and standard deviation of 2.4 × 10 −4 . Find the 95% Bayesian credible interval for the failure rate.

Solution: To obtain the 95% credible interval for the normal distribution, we find the 2.5th and 97.5th percents of the normal distribution representing uncertainty on the parameter.

λl = Φ −1 ( 0.025, 1 × 10 −3 ,  2.4 × 10 −4 ) = 5.30 × 10 −4 , λu = Φ −1 ( 0.975, 1 × 10 −3 ,  2.4 × 10 −4 ) = 1.47 × 10 −3. Therefore, 5.30 × 10 −4 < λˆ < 1.47 × 10 −3 . This is illustrated in the figure below.

4.6

HYPOTHESIS TESTING AND GOODNESS OF FIT

Another important aspect of reliability estimation is indicating how well a set of observed data fits a known distribution. One way to do this is to determine whether a hypothesis that the data originate from a known distribution is true. For this purpose, we need to perform a test that determines the adequacy of a fit by evaluating the difference (e.g., the distance) between the frequency of occurrence of an r.v. characterized by an observed sample and the expected frequencies obtained from the hypothesized distribution. For this purpose, the goodness of fit tests are used. It is necessary to calculate the expected frequencies of failures from the hypothesized distribution and measure the difference from the observed frequencies. Several

129

Basic Reliability Mathematics

methods exist to perform such a fit test. We discuss two of these methods in this book. For a more comprehensive discussion on goodness of fit tests the reader is referred to Rayner et al. (2009).

4.6.1

hyPothesis testing basics

Interval estimation and hypothesis testing may be viewed as mutually inverse procedures. Let us consider an r.v. X with a known pdf f ( x θ ). Using a random sample from this distribution, one can obtain a point estimate θˆ of parameter θ. Let θ have a hypothesized value of θ = θ 0 . Under these conditions, the following question can be raised: is the θˆ estimate compatible with the hypothesized value θ 0? In terms of statistical hypothesis testing, the statement θ = θ 0  is called the null hypothesis, which is denoted by H0. For the case considered, it is written as H 0: θ = θ 0 .

(4.22)

The null hypothesis is always tested against an alternative hypothesis, denoted by H1, which for the case considered might be the statement θ ≠ θ 0 , which is written as H1: θ ≠ θ 0 .

(4.23)

The null and alternative hypotheses are also classified as simple (or exact, when they specify exact parameter values) or composite (or inexact, when they specify an interval of parameter values). In the above equations, H0 is simple and H1 is composite. An example of a simple alternative hypothesis is H1: θ = θ * .

(4.24)

For testing statistical hypotheses, test statistics are used. Often the test statistic is the point estimator of the unknown distribution. Here, as in the case of the interval estimation, one must obtain the distribution of the test statistic used. Recall the example considered above. Let x1 , …, x n  be a random sample from the normal distribution, X ~ N ( µ , σ ) , in which μ is an unknown parameter and σ2 is assumed to be known. One must test the simple null hypothesis H 0: µ = µ * , 

(4.25)

H1: µ ≠ µ * .

(4.26)

against the composite alternative

As the test statistic in Equation 4.25, use the same sample mean, x , which has the σ   normal distribution, X ~ N  µ ,  . Having the value of the test statistic x , one can  n  construct the confidence interval using Equation 4.18 and see whether the value of

130

Reliability and Risk Analysis

μ* falls inside the interval. This is the test of the null hypothesis. If the confidence interval includes μ*, then we can conclude the null hypothesis is not rejected at the significance level γ . In terms of hypothesis testing, the confidence interval considered is called the acceptance region, and the upper and the lower limits of the acceptance region are called the critical values, and the significance level γ is called a probability of a Type I error. In deciding about whether to reject the null hypothesis, it is possible to commit these errors: • Reject H0 when it is true (Type I error). • Do not reject H0 when it is false (Type II error). The probability of the Type II error is designated by β. These situations are traditionally represented by Table 4.2. TABLE 4.2 Goodness of Fit Error Types State of Nature (True Situation) Decision Reject H0 Do not reject H0

H0 is True Type I error No error

H0 is False No error Type II error

Increasing the acceptance region decreases γ and simultaneously results in increasing β. The traditional approach to this problem is to keep the probability of Type I errors at a low level (0.01, 0.05, or 0.10) and to minimize the probability of Type II errors. The probability of not making a Type II error is called the power of the test.

4.6.2

chi-squared test

As the name implies, this test is based on a statistic that has an approximate χ2 distribution. To perform this test, an observed sample taken from the population representing an r.v. X must be split into k nonoverlapping (mutually exclusive) intervals. The lower limit for the first interval can be −∞, and the upper limit for the last interval can be +∞. The assumed (hypothesized) distribution model is then used to determine the probabilities pi  that the r.v. X would fall into each interval i (i = 1,…, k). By multiplying pi by the sample size n, we obtain the expected frequency for each interval i, denoted as ei (see Example 4.2). It is obvious that ei = npi . If the observed frequency for each interval i of the sample is denoted by oi , then the differences between ei  and oi can characterize the goodness of the fit. The χ2 test uses the test statistic χ 2 or W, which is defined as k

W = χ2 =

∑ (o −e e ) . i

i=1

i

i

2

(4.27)

Basic Reliability Mathematics

131

The χ2 statistic approximately follows the χ2 distribution. If the observed frequencies oi differ considerably from the expected frequencies ei , then W will be large, and the fit is considered poor. A good fit would lead to not rejecting the hypothesized distribution, whereas a poor fit leads to the rejection. It is important to note that the hypothesis can be rejected, but cannot be positively affirmed. Therefore, the hypothesis is either rejected or not rejected as opposed to accepted or not accepted. The chi-squared test has the following steps: Step 1: Choose a hypothesized distribution for the sample. Step 2: Select a significance level of the test, γ. Step 3: Calculate the value of the χ2 test statistic, W, using Equation 4.27. Step 4: Define the critical value (edge of rejection region) R ≥ χ (21−γ ) [ df ], where χ (21−γ ) [ df ] is 100(1–γ)% of the χ2 distribution with degrees of freedom df = k – m – 1, k is the number of intervals, and m is the number of parameters estimated from the sample. If the parameters of the distribution were estimated without using the sample, then m = 0. Step 5: If W > R, reject the hypothesized distribution; otherwise, do not reject the distribution. A few conditions must be met for the chi-squared test to be applicable. First, the data must be frequencies or counts, and they must be binned into mutually exclusive intervals. Generally, the value of the “expected frequency” cell should be greater than or equal to five in at least 80% of the cells. No cell should have an expected value of less than one. If these conditions are not met, either bin the data further, combine bins, or use another test where the assumptions are met. It is important at this point to specify the role of γ  in the χ2 test. Suppose that the 2 calculated value of W in Equation 4.27 exceeds the 95th percentile value, χ 0.95 [df], given in Table A.3. This indicates that chances are lower than 5% that the observed data are from the hypothesized distribution. Here, the model should be rejected; by not rejecting the model, one would make the Type II error. But if the calculated value 2 of W is smaller than χ 0.95 (·) , chances are greater than 5% that the observed data match the hypothesized distribution model. In this case, the model should not be rejected; by rejecting the model, one would make a Type I error. It is critical to recognize that not rejecting a hypothesis is not equivalent accepting it. One instructive step in χ2 testing is to compare the observed data with the expected frequencies to note which classes (intervals) contributed most to the value of W. This can sometimes help to understand the nature of the deviations. Example 4.5 The number of parts ordered per week by a maintenance department in a manufacturing plant is believed to follow a Poisson distribution. Use a χ 2 goodness of fit test to determine the adequacy of the Poisson distribution. Use the data found in the following table and a 0.05 significance level.

132 No. of Parts per Week, X

Reliability and Risk Analysis Observed Frequency, oi (No. of Weeks)

Expected Frequency, ei (No. of Weeks)

χ2 Statistic

( oi − ei ) 2 ei

0

18

15.783

0.311

1 2 3

18  8  5

18.818 11.219  4.459

0.036 0.923

4 5

 2  1

 1.329  0.317

52 weeks

52 weeks

Total:

}

6.1

0.588 W = 1.86

Solution: Since under the Poisson distribution model, events occur at a constant rate, then a natural estimate (and MLE) of μ is Number of parts used 62 µˆ = = = 1.192 parts/week. Number of weeks 52 From the Poisson distribution,

Pr( X = xi ) =

µe − µ . xi !

Using ρˆ = 1.192 , one obtains Pr(X = 0)=0.304. Therefore, ei = 0.304 ⋅ 52 = 15.783. The expected frequency for other rows is calculated in the same way. Since we estimated one parameter (Γ ) from the sample data, m = 1. Notice that in the table above, we have binned the expected frequencies for 3, 4, and 5 parts per week. You can see that the original values did not meet the criterion that most bins have an expected value greater than 5. Thus, these three adjacent bins were combined into one that met the criterion. Therefore, k = 4 bins. 2 Using Table A.3, we determine that R = χ 0.95 [6 − 4 − 1] = 3.84. Since (W = 1.86) 46,226 Total:

Observed Frequency, Oi 5 9 13 14 20 15 8 9 7 100

Expected Frequency, ei 5.446 7.124 11.750 15.916 17.704 16.172 12.131 7.473 6.077 99.792a

(oi − ei ) 2 ei 0.0365 0.4942 0.1329 0.2306 0.2978 0.0849 1.4066 0.3122 0.1401 W = 3.136

Note that the expected frequency column adds up to slightly less than 100. This is due to a small amount of density below 0 in the normal distribution; in this case it is negligible. In other cases, it may be necessary to use a truncated normal distribution.

a

Since both distribution parameters were estimated from the given sample, 2 m = 2. The critical χ 2 value of the statistic is, therefore, χ 0.95 [9 − 2 − 1] = 12.59. This is higher than the test statistic W = 3.0758; therefore, there is no reason to reject the hypothesis about the normal distribution at the 0.05 significance level. Visually

134

Reliability and Risk Analysis

assessing the plot of the observed frequencies vs. expected frequency further supports this conclusion.

4.6.3

kolMogorov-sMirnov (k-s) test

In the framework of the K-S test, the individual sample items are treated without clustering them into intervals. Similar to the χ 2 test, a hypothesized cdf is compared with its estimate known as the empirical cdf or sample cdf. A sample cdf is defined for an ordered sample x1 < x 2 <  < x n as  0,   i Sn ( x ) =  ,   n  1,

−∞ < x < x1 ,  xi ≤ x < x(i+1) ,  i = 1,2,…, n − 1,

(4.28)

xn ≤ x < ∞.

The K-S test statistic, D, measures the maximum difference between the sample cdf Sn ( x )and a hypothesized cdf, F(x). It is calculated as: D = max  F ( xi ) − Sn ( xi ) ,  F ( xi ) − Sn ( xi −1 )  . i

(4.29)

Similar to the χ2 test, the following steps compose the K-S test: Step 1: Choose a hypothesized cumulative distribution F(x) for the sample. Step 2: Select a specified significance level of the test, γ. Step 3: Define the rejection region R > Dn (γ ); critical values of the test statistic Dn (γ ) can be obtained from Table A.4. Step 4: If D > Dn (γ ), reject the hypothesized distribution and conclude that F(x) does not fit the data; otherwise, do not reject the hypothesis.

135

Basic Reliability Mathematics

Example 4.7 Time to failure of a group of electronic devices is measured in a life test. The observed failure times are 254, 586, 809, 862, 1381, 1923, 2,542, and 4,211 hours. Use a K-S test with γ = 0.05 to assess whether the exponential distribution with λ = 5 × 10 −4 is an adequate representation of this sample.

Solution: For γ = 0.05, D8 ( 0.05) = 0.454 . Thus, the rejection region is D > 0.454. We use Equation 4.29 to calculate the sample cdfs for Equation 4.29 as shown in the table −4 below. For an exponential distribution with λ = 5 × 10 −4, F ( t ) = 1 − e −5  ×  10 t . Using this, we calculate the fitted cdfs in the table below. The elements of Equation 4.29 are shown in the last two columns. As can be seen the maximum value is obtained for i = 7 as D = 0.156, shown in bold in the table below. Since ( D = 0.156 ) < 0.454 , we should not reject the hypothesized exponential distribution model with the given λ . Visually assessing the plot of the sample cdf vs. hypothesized cdf further supports this conclusion. Empirical cdf

Hypothesized cdf

Time to failure, ti

i

Sn ( ti )

Sn ( ti−−1 )

Fn ( ti )

Fn ( ti ) − Sn ( ti )

Fn ( ti ) − Sn ( ti−−1 )

254 586 809 862 1,381 1,923 2,542 4,211

1 2 3 4 5 6 7 8

0.125 0.250 0.375 0.500 0.625 0.750 0.875 1.000

0.000 0.125 0.250 0.375 0.500 0.625 0.750 0.875

0.119 0.254 0.333 0.350 0.499 0.618 0.719 0.878

0.006 0.004 0.042 0.150 0.126 0.132 0.156 0.122

0.119 0.129 0.083 0.025 0.001 0.007 0.031 0.003

4.7

LINEAR REGRESSION

Reliability engineering and risk assessment problems often require relationships among several r.v.s or between random and nonrandom variables. For example, time

136

Reliability and Risk Analysis

to failure of an electrical generator can depend on its age, environmental temperature, applied mechanical stresses and power capacity. Here, we can consider the time to failure as an r.v. Y, which is a function of the r.v.s X1  (age), X 2 (thermal and mechanical stresses), and X 3  (power capacity). In regression analysis we refer to Y as the dependent variable and to X1 , …, X j   as the j independent variables, explanatory variables, or features the latter is used (in machine learning). Note that the explanatory variables X i do not have to be independent of each other. Generally speaking, explanatory factors might be random or nonrandom variables whose values are known or chosen by the experimenter (in the case of the design of experiments (DoE)). The conditional expectation of Y for any X1 , …, X j , E (Y X1 , …, X j ) is known as the regression of Y on X1 , …, X j . In other words, regression analysis estimates the average value for the dependent variable corresponding to each value of the independent variable. When the regression of Y is a linear function with respect to the explanatory variables X1 , …, X j with values x1 , …, x j  it can be written in the form E (Y X 1 ,  X 2 , …, X n ) = β 0 + β1x1 + + β j x j .

(4.30)

The coefficients β = {β 0 , …,  β j } are called regression coefficients or model parameters. When the expectation of Y is nonrandom, the relationship Equation 4.30 is deterministic. The corresponding regression model for the r.v. Y ~ f ( y x ,  β , σ 2 ) , can be written in probabilistic form: Y = β 0 + β1x1 +  + β j x j + ε ,

(4.31)

where ε is the random model error, assumed to be independent (for all combinations of X considered) and distributed with mean E (ε ) = 0 and the finite variance σ2. One distribution that is often used to describe the model error is the normal distribution, known as the normal regression. The model parameters for the deterministic regression are usually obtained using an optimization approach. In this approach, the model parameters are found such that an objective function represented by a measure of total distance between the independent variables represented by values x1 , …, x j  and the regression model in Equation 4.31, known as a loss function, is minimized. An example of such a loss function for a single independent variable is

(

)

the square error loss function E β 0 ,  β1 , σ 2 =

n

∑ ( f ( y x ; β , β , σ ) − y ) , where i

i=1

0

1

2

2

i

n is the total number of data pairs ( yi ,  xi ) . To avoid any overfitting, there are additional terms known as penalty terms may also be added to the optimization. For more readings on deterministic regression using optimization techniques see Mendenhall and Sincich (2020). In the following, we will discuss a simple probabilistic solution to the estimation of the two-parameter linear regression models. For the simple linear regression model consider the simple deterministic relationship between Y and one explanatory factor X:

Basic Reliability Mathematics



137

138

Reliability and Risk Analysis

The estimate of the dependent variable variance, Var ( y ) = σ 2, can be found as

Var ( y ) = σ

∑ =

2

n

(yi − yˆi )2

i=1

n−2

,

(4.38)

where yˆi = βˆ0 + βˆ1xi , is predicted by the regression model values for the dependent variable and (n–2) is the number of degrees of freedom (and 2 is the number of the estimated parameters of the model). The estimate of the variance of Y from Equation 4.38 is called the residual variance and it is used as a measure of accuracy of model fit. The square root of Var(y) in Equation 4.38 is called the standard error of the estimate of y. As noted, the numerator in Equation 4.38 is called the square error loss function. For a more detailed discussion on the reliability applications of regression analysis, see Lawless (2002). It is worth mention that alternatively one may minimize x residuals instead of y using S ( β 0 ,  β1 ) =

n

∑ i=1

2

yi β 0    xi − β − β  ,  1 1

(4.39)

in which case the corresponding estimates of the regression parameters are

∑ (∑ y ) ,  βˆ = n∑ x y − ( ∑ x ) ( ∑ y ) yi2 −

n

2 i

2

1

i

i i

(4.40)

i

and

βˆ0 =

∑ y − βˆ ∑ x . i

n

i

1

n

(4.41)

Regression analysis is a major topic in machine learning. For more detailed methods and discussions refer to Bishop (2006). Example 4.8 An electronic device was tested under the elevated temperatures of 50°C, 60°C, and 70°C. The test results as times to failure (in hours) for a sample of 30 tested items are given below.

139

Basic Reliability Mathematics Time to Failure, t (hr)

ln(t)

50°C

60°C

70°C

50°C

60°C

70°C

1,950 3,418 4,750 5,090 7,588 10,890 11,601 15,288 19,024 22,700

607 644 675 758 1,047 1,330 1,369 1,884 2,068 2,931

44 53 82 88 123 189 204 243 317 322 E [ln(t)]

7.5756 8.1368 8.4659 8.5350 8.9343 9.2956 9.3588 9.6348 9.8535 10.0301 8.9820

6.4085 6.4677 6.5147 6.6307 6.9537 7.1929 7.2218 7.5412 7.6343 7.9831 7.0549

3.7842 3.9703 4.4067 4.4773 4.8122 5.2417 5.3181 5.4931 5.7589 5.7746 4.9037

This is an example of accelerated life testing with the logarithm of time to failure, t, follows the normal distribution with the mean given by the Arrhenius model, E ( ln t ) = A +

B ,  T

where T is the absolute temperature in Kelvin. Find estimates of parameters A and B.

Solution: The equation above can be transformed to the simple linear regression (Equation 4.32), y = β 0 + β1 x , using y = ln ( t ), x = 1/T, β 0 = A, and β1 = B. Accordingly, from the data in the table. y = E[ln(t)]

1 T

T(°C)

T(K)

8.9820

50

323.15

0.0031

7.0549 4.9037

60 70

333.15 343.15

0.0030 0.0029

x=

Using the transformed data above and the least squares estimates from Equation 4.37, we find the estimates of parameters A and B as Aˆ = −54.19, Bˆ = 20,392 K.

140

Reliability and Risk Analysis

4.8 EXERCISES

4

7

8

12

19

27

50

65

66

69

71

73

75

91

107

115

142

166

184

192



493 1,199 1,570 1,931 2,367

510 1,210 1,584 1,961 2,408



849 1,214 1,632 1,982 2,449

901 1,325 1,639 2,020 2,515

955 1,361 1,689 2,113 2,524

967 1,457 1,859 2,147 2,529

1,035 1,464 1,860 2,153 2,567

1,068 1,473 1,872 2,168 2,789

1,078 1,539 1,876 2,259 2,832

1,135 1,541 1,893 2,267 2,882

141

Basic Reliability Mathematics Time Interval (hour) 0–100

Observed Frequency 3

100–200 200–300 300–400 400–500 500–600 600–700 Total:

5 9 11 6 4 2 40



Year

Times Used

Unwanted Pregnancies (X)

1

8,200

19

2 3 4 5

10,100 2,120 6,120 18,130

18 1 9 30



142

Reliability and Risk Analysis



Time Interval (hour) Obs. Freq. 45–55

7

55–65 65–75 75–85 85–95

18 35 28 12



Digit

0

1

2

3

4

5

6

7

8

9

Frequency

4

8

8

4

10

3

2

2

4

5



Time Interval (hour) 0–2 2–6 6–10 10–15 15–25 25–100



Obs. Freq. 6 12 7 6 7 2

143

Basic Reliability Mathematics Time Range (hour) 0 4 24 72 300

Frequency

4

17

24 72 300 5,400

41 12 7 9



Interval (days)

Number of Failures

0 < T < 50 50 < T < 75 75 < T < 100 100 < T < 150 150 < T < 200

6 10 8 9 7



Interval (hour)

Observed Failures

0–100

16

100–200 200–300 300–400 400–500

17 22 22 23



# of Stoppages (x)

Frequency (days)

0

16

1 2 3

24 15 5

144

Reliability and Risk Analysis

Event

1

2

3

4

5

6

7

8

9

10

Time to Failure (hour)

10.3

12.4

13.7

13.9

14.1

14.2

14.4

15.0

15.9

16.1



X

1

2

3

4

5

6

7

Y

9

8

10

12

11

13

14

REFERENCES Bishop, C. M., Pattern Recognition and Machine Learning. Vol. 4, New York: Springer, 2006. Lawless, J. F., Statistical Models and Methods for Lifetime Data. 2nd Edition, Wiley, New York, 2002. Mendenhall, W. and T. T. Sincich, A Second Course in Statistics: Regression Analysis. 8th edition, Pearson, Hoboken, NJ, 2020. Rayner, J. C. W., O. Thas, and D. J. Best, Smooth Tests of Goodness of Fit. 2nd edition, John Wiley & Sons (Asia), Singapore, 2009.

5

Reliability Data Analysis and Model Selection

In Chapter 3, we discussed several common distribution models useful for reliability analysis of items. It is necessary at this point to discuss how field and reliability test data can support the selection and estimation of the parameters of a probability distribution model for reliability analysis. In this chapter, we first describe the types of data used in reliability model development, then we discuss the common procedures for selecting and estimating the model parameters using data and information. These procedures can be divided into two groups: nonparametric methods (that do not need a particular distribution function) and parametric methods (that are based on a selected distribution function). We discuss each in more detail. We will also discuss the methods of accounting for some uncertainties associated with model selection and estimated model parameters.

5.1 CONTEXT AND TYPES OF DATA 5.1.1

tyPes of field and test data

Data and information are critical components of reliability analysis. Typically, reliability data in the form of failure reports, warranty returns, complaints, surveys and tests are gathered and used in reliability model developments. Reliability data are often limited, incomplete, and may be biased and subjective. Regardless, it is essential to gather and use this data despite the limitations, and appropriate mathematical techniques have been designed to enable this. In many cases, it is straightforward to adapt current field data collection programs to include information that is directly applicable to reliability reporting Most of the reliability data and information come from one or more of the following sources: • • • • • • • •

Reliability tests (prototype or production) Environmental tests Reliability growth tests Production returns including failure analysis Customer returns Surveillance, maintenances, and field service repairs Generic databases Customer/user surveys.

In life testing for reliability estimation, a sample of components from a sample population of such components is tested replicating the environment in which the DOI: 10.1201/9781003307495-5

145

146

Reliability and Risk Analysis

components are expected to work, and their times to failure (or censored times) are recorded. In general, two major types of tests are performed. The first is testing with replacement of the failed items, and the second is testing without replacement of the failed items. The test with replacement is sometimes called monitored testing. Reliability data obtained from these sources can be divided into two types: complete data and censored data. In the following each type is discussed further.

5.1.2

coMPlete data

When ti , the exact time or cycle of failure for a specific failure mode is available for all items ( i = 1,…, n ) , the data are complete. For example, consider an accelerated reliability test where the items under test are continuously monitored and when a failure occurs it is immediately recorded. If the test is run until all the items fail, and those times are recorded, this type of data is a complete failure data set.

5.1.3

censored data

In some field data gathering and testing, we reach the end of a test or our observation and some items may have not failed, or we detect failure events that are of no interest to us (e.g., the failure is not due to the failure mode of interest) or the exact time of failure is not known. This type of data is called censored data. There are three types of censored data: right censored (also called suspended data), interval censored, and left censored. In the following, we will further elaborate on each type of censoring. 5.1.3.1 Left, Right, and Interval Censoring Censoring is a form of missing data characterization. It describes cases when complete failures do not occur, or the precise time of failure is not observed because, for example, a reliability test was terminated before some items failed. Censoring also includes cases when the failure event was of no interest (e.g., the failure was due to observing an unrelated failure mode), or the item was removed from the observation before failing. Censoring is very common in reliability data analysis. Let n be the number of items in a sample and assume that all items of the sample are tested simultaneously. If in the test duration, t test, only r   of the n items have failed, the failure times are known, and when the failed items are not replaced, the sample is singly censored on the right at t test. In this case, the only information we have about the n – r non-failed (censored) items is that their failure times are greater than t test, if they were allowed to continue working. Formally, an observation is right censored at t test, if the exact time of failure is not known, but it is known that it is greater than or equal to t test. If the failure time for a failed item is not known but is known to be less than a given value, the failure time is called left censored. This type of censoring is not very common in reliability data collection, and so it is not discussed in detail, but will be introduced later in this chapter when the topic of likelihood functions of the reliability data including censored data is introduced. If the only information available is that an item is failed in an interval (e.g., between successive inspections), the respective data are grouped or interval censored. Interval

Reliability Data Analysis and Model Selection

147

censoring commonly occurs when periodic data observation and collection (reliability test inspections) are used to assess if a failure event of interest has occurred. A special kind of censoring, right-truncated censoring, refers to cases where failure times are truncated if they occurred after a set time  tU . Similarly, left-truncated censoring refers to cases where failure data are truncated for those failures that occurred before a set time   t L . Finally, interval-truncated censoring refers to cases where a failure occurred at time ti should be truncated on both sides if the observations were between t L and tU . Again, these conditional censoring methods are not very common in reliability analysis. It is important to understand the way—or mechanism by which—censored data are obtained. The basic discrimination is associated with random and nonrandom censoring, the simplest cases of which are discussed below. 5.1.3.2 Type I Censoring Consider the situation of right censoring. If the test is terminated at a given nonrandom time, t test, then r, the number of failures observed during the test period, is an r.v. These censored data are Type I or time right-singly-censored data, and the corresponding test is sometimes called time terminated. For the general case, Type I censoring is considered under the following scheme of observations. In the timeterminated life test, n items are placed on a test and the test is terminated after a predetermined time has elapsed. The number of components that failed during the test time and the corresponding time to failure of each component are recorded. Let each item in a sample of n items be observed (during reliability testing or in the field) during different periods of time L1 ,…,  Ln . The time to failure of an individual item, ti , is a distinct value if it is less than the corresponding period, that is, if   ti < Li . Otherwise, ti is the time to censoring, which indicates that the time to failure of the ith item is greater than Li . This is the case of Type I multiply-censored data; the singly-censored case considered above is its particular case, when L1 =  = Ln = t test . Type I multiply-censored data are common in reliability testing. For example, a test may start with a sample size of n, but at some given times L1 ,…,  Lk   (where k < n ), a prescribed number of items are removed from (or placed on) the test. Another source of censoring typically found in reliability data analysis when there are several failure modes whose times to failure must be estimated separately. The times to failure due to each failure mode are considered r.v.s having different distributions, while the whole system is called a competing risks (or series) system. An example of this type of censoring is collecting failure data during the warranty service of a household product. The manufacturers are interested in reliability estimations and any defects in this product. There could be failures attributed to manufacturing, but failures could also be user-induced or caused by poor maintenance and repair. For estimating manufacturer-related failures, failures due improper use of the product or maintenance-related failures must be treated as time censored. Note that the situation might be opposite. For example, an organization (say, a regulatory agency) might be studying the safety performance of the product, so they would need to estimate the rates of any user-induced failures. In this case, the failures attributed to manufacturing causes should be censored. This differentiation in censored data is important in reliability database developments.

148

Reliability and Risk Analysis

5.1.3.3 Type II Censoring A test may also be terminated when a nonrandom number of failures (say, r), specified in advance, occurs. Here, the duration of the test is an r.v. This situation is known as Type II right censoring and the corresponding test is called a failure-terminated test. Under Type II censoring, only the r smallest times to failure t1 <  < tr out of the sample of n times to failure are known times to failure. The times to failure ti   (i = 1, …, r) are identically distributed r.v.s (as in the previous case of the Type I censoring). In the failure-terminated life tests, n items are placed on test and the test is terminated when a predetermined number of component failures have occurred. The time to failure of each failed component, including the time of the last failure, is recorded. Type I and Type II life tests can be performed with replacement or without replacement.

5.2 RELIABILITY DATA SOURCES Due to the lack of observed data, component reliability analysis may require the use of generic (base) failure data representing an average value over many organizations or industries adjusted for the factors that influence the failure rate for the component under analysis. Generally, these classes of influencing factors are considered. 1. Environmental factors: These factors affect the failure rate due to extreme mechanical, electrical, nuclear, and chemical environments. For example, a high-vibration environment would lead to high stresses that promote failure of components. 2. Design and manufacturing factors: These factors affect the failure rate due to the quality of material used or workmanship, material composition, functional requirements, geometry, and complexity. 3. Operating factors: These factors affect the failure rate due to the applied stresses resulting from operation, testing, repair, maintenance practices, and so on. To a lesser extent, the age factor is used to correct for the infant and wear-out periods, and the original factor is used to correct for the accuracy and dependability of the data source (generic data). For example, obtaining data from observed failure records as opposed to expert judgment may affect the failure rate dependability. The failure rate can be represented as

λa = λ g K E K D KO …,

(5.1)

where λa is the actual (or adjusted) failure rate and λ g   is the generic (base) failure rate, and K E ,  K D , KO , … are correction (adjusting) factors for the environment, design, and operation, respectively. Other multiplicative factors can be added to account for additional factors such as materials, installation, construction, and manufacturing. It is possible to subdivide each of the correction factors into their contributing subfunctions accordingly. For example, K E = f ( ka , kb ,…), when ka and kb are factors such as vibration level, moisture, and pH level. These factors vary for different types of components.

Reliability Data Analysis and Model Selection

149

This concept is used in the procedures specified in government contracts for determining the actual failure rate of electronic components. The procedure is summarized in MIL-HDBK-217. In this procedure, a base failure rate of the component is obtained from a table, and then they are multiplied by the applicable adjusting factors for each type of component. For example, the actual failure rate of a tantalum electrolytic capacitor is given by

λc = λb ( π E ·π SR ·π Q ·π CV ) ,

(5.2)

where λc is the adjusted (actual) component failure rate and λb is the base (or generic) failure rate, and the π factors are correction factors for the environment (E), series resistance (SR), quality (Q), and capacitance (CV). Values of λb and the correction factors are given in MIL-HDBK-217 for many types of electronic components. Generally, λb for electronic items is obtained from the Arrhenius empirical model: E

λ b = Ae kT ,

(5.3)

where E = activation energy for the process (treated as a model parameter), T = absolute temperature (K), k = 1.38 × 10 −23  JK −1 is the Boltzmann constant, and A = a constant. The readers are referred to the software 217Plus (2015) for the most updated generic data on electronics components. Further, readers interested in the extension of the physics-based generic failure data for electronic components should refer to Salemi et al. (2008). In MIL-HDBK-217 the Arrhenius model forms the basis for a large portion of electronic components subject to temperature degradation. However, care must be applied in using this database, especially because the data in this handbook are derived from repairable systems (and hence, apply to such systems) and assume constant failure rates. Also, application of the various correction factors can drastically affect the actual failure rates. Therefore, proper care must be applied to ensure correct use of the factors and to verify the adequacy of the factors. Also, the appropriateness of the Arrhenius model has been debated many times in the literature. Many generic sources of data are available for a wide range of items. Among them are IEEE-500, CCPS Guidelines for Process Equipment Data, IEC/TR 62380 Reliability data handbook, and nuclear power plant Probabilistic Risk Assessment data sources such as NUREG/CR-4550 (Ericson et al., 1990) and NUREG/CR-6928 (Eide et al., 2007). For illustration, Appendix B shows a set of data from the 2020 update to NUREG/CR-6928 (Ma et al., 2021). The following are some key reliability data sources: • MIL-HDBK-217F - Reliability Prediction of Electronic Equipment: http://everyspec.com/MIL-HDBK/MIL-HDBK-0200-0299/MIL-HDBK-217F_14591/ • EPRD - Electronic Parts Reliability Data (RIAC): https://www.quanterion. com/?s=riac • NPRD-95 Non-electronic Parts Reliability Data (RIAC): https:// www.quanter ion.com /product /publications/nonelectronic-pa r tsreliability-data-publication-nprd-2016/

150

Reliability and Risk Analysis

• FMD-97 Failure Mode/Mechanism Distributions (RIAC): https://www. quanterion.com/?s=riac • FARADIP (FAilure RAte Data In Perspective) https://www.m2k.com/FARADI P(FAilureRAteDataInPerspective) • SR-332 Reliability Prediction for Electronic Equipment (Telcordia Technologies): http://www.ericsson.com/ourportfolio/telcordia_landingpage • FIDES (mainly electronic components): http://www.fides-reliability.org/ • IEC/TR 62380 Reliability Data Handbook - Universal model for reliability prediction of electronics components, PCBs and equipment: http://www. fides-reliability.org/ • EiReDA - European Industry Reliability Data Mainly components in nuclear power plants • OREDA - Offshore Reliability Data: Topside and subsea equipment for offshore oil and gas production: http://www.oreda.com/ • Handbook of Reliability Prediction Procedures for Mechanical Equipment Mechanical equipment - military applications • T-Book (Reliability Data of Components in Nordic Nuclear Power Plants, ISBN 91-631-0426-1) • Reliability Data for Control and Safety Systems - PDS Data Handbook Sensors, detectors, valves & control logic: http://www.sintef.no/Projectweb/ PDS-Main-Page/PDS-Handbooks/ • Safety Equipment Reliability Handbook (exida): Safety equipment (sensors, logic items, actuators): http://www.exida.com/ • WellMaster (ExproSoft): Components in oil wells: http://www.exprosoft.com/ • SubseaMaster (ExproSoft): Components in subsea oil/gas production systems: http://www.exprosoft.com/ • PERD - Process Equipment Reliability Data(AIChE): Process equipment: http:// www.aiche.org/ccps/resources/process-equipment-reliability-database-perd • GIDEP (Government-Industry Data Exchange Program): http://www.gidep. org/ • CCPS Guidelines for Process Equipment Reliability Data, AIChE, 1989 Process equipment • IEEE Std. 500–1984: IEEE Guide to the Collection and Presentation of Electrical, Electronic, Sensing Component, and Mechanical Equipment Reliability Data for Nuclear Power Generating Stations

5.3 5.3.1

NONPARAMETRIC AND PLOTTING METHODS FOR RELIABILITY FUNCTIONS nonParaMetric Procedures for reliability functions

The nonparametric approach, in principle, attempts to directly estimate the reliability characteristic of an item (e.g., the pdf, reliability, and hazard rates) from a sample. The shape of these functions, however, is often used as an indication of the most

151

Reliability Data Analysis and Model Selection

appropriate parametric distribution representation. Thus, such procedures can be considered tools for exploratory (preliminary) data analysis. It is important to mention a key assumption that each failure time must be considered an identical sample observation and independent of the other item failure times, thus meeting the identical and independently distributed (i.i.d.) criterion used in statistical analysis. With failure data from a repairable component, these methods can only be used if the item is assumed to be as good as new following repair or maintenance. Under the as-good-as-new assumption, each failure time can be considered an identical sample observation independent of the previously observed failure times, meeting the i.i.d. criterion used in statistical analysis. Therefore, n observed times to failure of such a repairable component is equivalent to putting n independent new components under life test 5.3.1.1

Nonparametric Component Reliability Estimation Using Small Samples Suppose n times to failure makes a small sample (e.g., n < 25). Let the data be ordered such that t1 ≤  ≤ tn . Blom (1958) introduced the nonparametric estimators of the hazard rate function, reliability function and pdf as: hˆ ( ti ) =

1 , i = 1,…, n − 1, ( n − i + 0.625)( ti +1 − ti )

(5.4)

n − i + 0.625 Rˆ ( ti ) = , i = 1,…, n, n + 0.25

(5.5)

and fˆ ( ti ) =

1

( n + 0.25)( ti +1 − ti )

, i = 1,…, n − 1.

(5.6)

Although there are other estimators besides those above, Kimball (1960) concludes that estimators in Equations 5.4–5.6 have good properties and recommends their use. One should keep in mind that 0.625 and 0.25 in Equations 5.4–5.6, which are sometimes called the Kimball plotting positions, are correction terms of a minor importance, which result in a small bias and a small mean square error. Other popular estimators of R ( ti ) include the mean rank ( n − i + 1) / ( n + 1) and the median rank ( n − i + 0.7 )/( n + 0.4 ). See Kapur and Lamberson (1977) for more detail. Example 5.1 A high-pressure pump in a process plant has the failure times ti (in thousands of hours, shown in the following table). Use a nonparametric estimator and the plotting positions in Equations 5.4–5.6 to calculate hˆ ( ti ) , Rˆ ( ti ) ,  fˆ ( ti ) . Plot hˆ ( ti ) , and discuss the results.

152

Reliability and Risk Analysis

Solution: i

ti

ti+1 − ti

hˆ ( ti )

Rˆ ( ti )

fˆ ( ti )

1

0.20

0.60

0.25

0.91

0.23

2 3 4 5 6 7

0.80 1.10 1.51 1.83 2.52 2.98

0.30 0.41 0.32 0.69 0.46 —

0.59 0.53 0.86 0.55 1.34 —

0.78 0.64 0.50 0.36 0.22 0.09

0.46 0.34 0.43 0.20 0.30 —

From the above histogram, one can conclude that the hazard rate is somewhat constant over the operating period of the component, with an increase toward the end. However, as a point of caution, although a constant hazard rate might be concluded, several other tests and additional observations may be needed to support this conclusion. Additionally, the histogram is only a representative of the case under study. An extension of the result to future times or other cases (e.g., other high-pressure pumps) may not be accurate.

5.3.1.2 Nonparametric Component Reliability Estimation Using Large Samples Suppose n times to failure make a large sample. Suppose further that the times to failure in the sample are grouped into several equal increments, Δt. According to the definition of reliability, a nonparametric estimate of the reliability function is N (t ) Rˆ ( ti ) = s i , N

(5.7)

where N s ( ti ) represents the number of surviving components in the interval starting in  ti . This approach is sometimes called the Nelson-Aalen approach. Note that the estimator (Equation 5.7) is compatible with the empirical (sample) cdf (Equation 4.28) introduced in Chapter 4. Time ti is usually taken to be the lower endpoint of each interval, although this may differ among practitioners. Similarly, the pdf is estimated by N (t ) fˆ ( ti ) = f i , N∆t

(5.8)

153

Reliability Data Analysis and Model Selection

where N f ( ti ) is the number of failures observed in the interval (ti ,  ti + ∆t ). Finally, using Equations 5.7 and 5.8, one obtains N f ( ti ) hˆ ( ti ) = . N s ( ti ) ∆t

(5.9)

For i = 1, N s ( ti ) = N , and for i > 1, N s ( ti ) = N s ( ti −1 ) − N f ( ti ). Equation 5.9 gives an estimate of average hazard rate during the interval (ti ,  ti + ∆t ). When N s ( ti ) → ∞ and ∆t → 0, the estimate from Equation 5.9 approaches the true hazard rate h(t). N (t ) In Equation 5.9, f i is the estimate of probability that the component will fail N s ( ti ) in the interval (ti ,  ti + ∆t ), since N s ( ti ) represents the number of components functioning at ti . Dividing this quantity by Δt, the estimate of hazard rate (probability of failure per surviving item entering an interval Δt) is obtained. It should be noted that the accuracy of this estimate depends on Δt. Therefore, if smaller Δt values are used, we would, theoretically, expect to obtain a better estimation. However, the drawback of using smaller Δt values is the decrease in the amount of data for each interval to estimate Rˆ ( ti ), fˆ ( ti ) and hˆ ( ti ). Therefore, selecting Δt requires consideration of both opposing factors. Example 5.2 Times to failure for an electrical device are obtained during three stages of the component’s life. The first stage represents the infant mortality of the component; the second stage represents chance failures; and the third stage represents the wear-out period. Plot the hazard rate for this component using the data provided below.

Solution: Use the Nelson-Aalen approach to calculate the empirical hazard rate, reliability, and pdf. Given Data Interval (ti)

Calculated Data Frequency

Beginning

End

N f ( ti )

0 20 40 60 80 100 120

20 40 60 80 100 120 >120

79 37 15 6 2 1 10 Total 150

N s ( ti )

hˆ ( ti )

Rˆ ( ti )

fˆ ( ti )

0.02633 0.02606 0.02206 0.01579 0.00769 0.00455 0.05000

1.00000 0.47333 0.22667 0.12667 0.08667 0.07333 0.06667

0.02633 0.01233 0.00500 0.00200 0.00067 0.00033 0.00333

Infant Mortality Stage 150 71 34 19 13 11 10

(Continued)

154

Reliability and Risk Analysis Given Data Interval (ti)

Beginning

0 2,000 4,000 6,000 8,000 10,000

Frequency End

N f ( ti )

N s ( ti )

hˆ ( ti )

Rˆ ( ti )

fˆ ( ti )

1.00000 0.57800 0.29400 0.16000 0.10400 0.06200

0.00021 0.00014 0.00007 0.00003 0.00002 0.00003

0.00113 0.00278 0.00573 0.01000

1.00000 0.88667 0.64000 0.27333

0.00113 0.00247 0.00367 0.00273

Chance failures

Wear-out

Chance Failure Stage 211 500 0.00021 142 289 0.00025 67 147 0.00023 28 80 0.00018 21 52 0.00020 31 31 0.00050 Total 500

2,000 4,000 6,000 8,000 10,000 >10,000

0 100 200 300

Calculated Data

100 200 300 >300

34 74 110 82 Total 300

0.03

Wear-out Stage 300 266 192 82

Infant mortality

h (ti) (hr-1)

0.025 0.02 0.015 0.01 0.005 0

0

20

40

60

80

100

0

2000 6000 8000 10000 0 ti (hr)

100 200 300

The graph above plots the estimated hazard rate functions for the three observation periods. Note that the three periods, each having a different scale, are combined on the same x-axis for demonstration purposes. The chance failure region represents most of the device’s life.

5.3.2 Probability distribution Plotting using life data Probability plotting is a simple nonparametric graphical method of displaying and analyzing observed data. The data are plotted on properly scaled probability papers or coordinates that transform a life cdf to plot as a straight line. Each type of distribution has its own probability plot scales. If a set of data is hypothesized to originate from a known distribution, the graph can be used to conclude whether the hypothesis might be

Reliability Data Analysis and Model Selection

155

rejected or not. From the plotted line, one can also roughly estimate the parameters of the hypothesized distribution. Probability plotting is often used in reliability analysis as an approximate first step to test the appropriateness of using known distributions to model a set of observed data. This method is used because it provides simple, visual representation of the data. Because of its approximate nature, the plotting method should be used with care and preferably as an exploratory data analysis procedure. With modern computing techniques, the probability paper itself has become obsolete. However, the term “probability paper” is still used to refer to the graphical method of parameter estimation. It takes a simple linear regression method to find an algebraic expression for the model. We will briefly discuss the scaled probability coordinates for the basic distributions considered in this book. 5.3.2.1 Exponential Distribution Probability Plotting Taking the logarithm of both sides of the expression for the reliability function of the exponential distribution, one obtains lnR ( t ) = − λ t.

(5.10)

If R ( t ) is plotted as a function of time, t, on semi-logarithmic scaled coordinates, according to Equation 5.10, the resulting plot will be a straight line with the slope of (−λ) and a y-intercept of 0. Consider the following n times to failure observed from a life test. Note that the data must be ordered such that: t1 ≤  ≤ tn . According to Equation 5.10, a nonparametric estimate of the reliability R ( ti ) can be made for each ti . A crude, nonparameti ric estimate of R ( ti ) is  1 −  (recall Equations 5.7 and 4.28). However, as noted in  n Section 5.3.1.1, the Kimball estimator in Equation 5.5 provides a better estimate for R(t) for the Weibull distribution. Graphically, on the semi-logarithmic scale the y-axis shows R ( ti ) and the x-axis shows ti . The resulting points should reasonably fall on a straight line if these data can be described by the exponential distribution. Since the slope of ln  R ( t )  vs. t is negative, it is often more convenient to plot ln (1/ R ( t )) vs. t in which the slope is positive. It is also possible to estimate the MTTF from the plotted line. For this purpose, at the level of R = 0.368 (or 1/R = e ≈ 2.718), a line parallel to the x-axis is drawn. At the intersection of this line and the fitted line, another line vertical to the x-axis is drawn. The value of t on the x-axis is an estimate of MTTF, and its inverse is λˆ . Alternatively, a simple linear regression of ln 1/ R ( t ) vs. t can be performed to obtain the parameter λ and data scatter around the fitted line (i.e., a measure of model error). Since the exponential distribution is a particular case of the Weibull distribution, the Weibull plotting paper may also be used for the exponential distribution.

(

)

Example 5.3 Nine times to failure of a diesel generator are recorded as 31.3, 45.9, 78.3, 22.1, 2.3, 4.8, 8.1, 11.3, and 17.3 days. If the diesel generator is restored to as good as new after each failure, determine whether the data represent the exponential distribution. That is, find λˆ and Rˆ (193 hours).

156

Reliability and Risk Analysis

Solution: First arrange the data in increasing order and then calculate the corresponding Rˆ ( ti ) . i

n − i + 0.625 Rˆ ( ti ) = n + 0.25

T

 1  ln    Rˆ ( ti ) 

1

2.3

0.93

0.07

2 3 4 5 6 7 8 9

4.8 8.1 11.3 17.3 22.1 31.3 45.9 78.3

0.82 0.72 0.61 0.50 0.39 0.28 0.18 0.07

0.19 0.33 0.50 0.69 0.94 1.26 1.74 2.69

The figure below shows a plot of the above data on semi-logarithmic paper. Using linear regression and setting the y-intercept to 0 produces an estimated parameter of λˆ = 0.0364 failures/day. Similarly, implementing the least squares regression approach from Chapter 4 produces a parameter of λˆ = 0.0346 failures/ day and a small y-intercept of 0.0848. Using the estimated parameter to estimate reliability at 52 days, we get Rˆ ( 52 ) = e −0.0364 ×52 = 0.15.

157

Reliability Data Analysis and Model Selection

5.3.2.2 Weibull Distribution Probability Plotting One can transform the Weibull cdf to a linear form by taking the logarithm of both sides twice: β  t

  1 = e α  , R (t )

(5.11)

  1  ln  ln    = β lnt − β ln α .   R (t )  

This linear relationship (in lnt) provides the basis for the Weibull plots or regression   1  analysis. The dependent term ln  ln    plots as a straight line against the inde  R (t )   pendent variable ln t with slope   β and y-intercept ( − β ln α ) . Accordingly, the values of the Weibull parameters α and β can be obtained from the y-intercept and the slope of the graph or by using a regression approach to estimate these parameters. As was mentioned earlier, several nonparametric estimators of R(t) can be used to   1  determine ln  ln   , including the recommended Kimball estimator (Equation   R (t)   5.5). The corresponding plotting or linear regression procedure is straightforward. The degree to which the plotted data or the regression analysis follows a straight line determines the conformance of the data to the Weibull distribution. If the data points approximately fall on a straight line, the Weibull distribution is a reasonable fit, and the shape parameter and the scale parameter β can be roughly estimated. Similar to the exponential distribution, it is straightforward to create a plot or perform least squares estimation to obtain α and β by using the linearized form of the Weibull cdf. Example 5.4 Time to failure of a device is assumed to follow the Weibull distribution. Ten of these devices undergo a reliability test resulting in complete failures. The times to failure (in hours) are 89, 132, 202, 263, 321, 362, 421, 473, 575, and 663. If the Weibull distribution is the proposed model for these data, what are the parameters of this distribution? What is the reliability of the device at 1,000 hours?

Solution: i

1

2

3

4

5

6

7

8

9

10

ti

89

132

202

263

321

362

421

473

575

663

ln(ti )

4.49

4.88

5.31

5.57

5.77

5.89

6.04

6.16

6.35

6.50

n − i + 0.625 Rˆ ( ti ) = n + 0.25

0.939

0.841

0.744

0.646

0.549

0.451

0.354

0.256

0.159 0.061

  1  ln  ln     Rˆ ( ti )    

−2.77 −1.76 −1.22 −0.83 −0.51 −0.23 0.04

0.31

0.61

1.03

158

Reliability and Risk Analysis

Based on the plot, the fit is reasonably good. The graphical estimate of β (the slope of the line) is approximately 1.78, and the estimate of α is approximately 402 hours. Therefore, at 1,000 hours, the reliability is 0.006 [i.e., R(t = 1.000) = 0.006]. In cases where the data do not fall on a straight line but are concave or convex in shape, it is possible to find a location parameter γ (i.e., to try using the threeparameter Weibull distribution (Equation 3.28)) that might straighten out the data fit. For this procedure, see Nelson (2003). If the failure data are grouped, the class midpoints ti′ (rather than ti ) should be used for plotting, where ti′ = ( ti −1 + ti ) / 2. One can also use class endpoints instead of midpoints. Recent studies suggest that the Weibull parameters obtained by using class endpoints in the plot or regression match better with those of the MLE method.

5.3.2.3 Normal and Lognormal Distribution Probability Plotting As before, properly scaled probability papers can be used for both normal and lognormal plots. In both cases, F ( ti ) is estimated by an appropriate nonparametric estimator. For the normal distribution, ti   is plotted on the x-axis, and for the lognormal distribution ln ( ti )  is plotted. It is easy to show that the normal cdf can be linearized using the following transformation: Φ −1  F ( ti ) =

1 µ ti − , σ σ

(5.12)

where Φ −1 (·) is the inverse of the standard normal cdf. In the case of lognormal distribution, ti   in Equation 5.12 is replaced by ln ( ti ). Some lognormal papers are logarithmic on the x-axis, in which case ti can be directly expressed. If the plotted data fall on a straight line, a normal or lognormal distribution might be the appropriate life model.

159

Reliability Data Analysis and Model Selection

Estimating the parameters from the plot follows a similar procedure to the other distributions we have discussed so far. Using linear regression, the parameter σ is the inverse of the line’s slope, and µ is the y-intercept times σ . Another way to estimate µ is to use the x-intercept. Example 5.5 The time it takes for a thermocouple to drift upward or downward to an unacceptable level is measured and recorded (see the following table, with times in months). Determine whether the time to failure (i.e., drifting to an unacceptable level) can be modeled by a normal distribution. i − 0.375 Fˆ ( ti ) = n + 0.25

i

ti

1

11.2

0.044

2 3 4 5 6 7

12.8 14.4 15.1 16.2 16.3 17.0

0.114 0.184 0.254 0.325 0.395 0.465

(

)

(

i

ti

i − 0.375 Fˆ ( ti ) = n + 0.25

Φ −1 Fˆ ( ti )

−1.71

8

17.2

0.535

0.09

−1.21 −0.90 −0.66 −0.45 −0.27 −0.09

9 10 11 12 13 14

18.1 18.9 19.3 20.0 21.8 22.7

0.605 0.675 0.746 0.816 0.886 0.956

0.27 0.45 0.66 0.90 1.21 1.71

Φ −1 Fˆ ( ti )

)

Solution:

The resulting plot shows that the data reasonably follows the normal distribution, with µ = 17.21 months and σ = 3.42 months.

160

Reliability and Risk Analysis

Example 5.6 Five components undergo low-cycle fatigue crack tests. The cycles to failure for the five tests are: 363, 1,115, 1,982, 4,241, and 9,738. Determine if the lognormal distribution is a candidate model for this data. If it is, estimate the parameters of the lognormal distribution.

Solution: i − 0.375 Fˆ ( ti ) = n + 0.25

(

Φ −1 Fˆ ( ti )

i

ti (cycles)

1

363

5.89

0.119

−1.18

2 3 4 5

1,115 1,982 4,241 9,738

7.02 7.59 8.35 9.18

0.310 0.500 0.690 0.881

−0.50 0.00 0.50 1.18

In (ti)

)

From the probability plot above µ = 7.61 and σ = 1.39 . The mean and standard deviation of the unacceptable drift time can be estimated from the lognormal distribution parameters by using the properties of the lognormal distribution in Table 2.9). The mean, E ( ti ) = 5,302 cycles and the stan-

dard deviation is StDev ( ti ) = 12,913 cycles. Data appear to have a large scatter around the lognormal fit, and consideration and evaluation of fits to other distribution models is prudent.

5.4

MAXIMUM LIKELIHOOD ESTIMATION OF RELIABILITY DISTRIBUTION PARAMETERS

This section deals with the maximum likelihood estimation (MLE) method for estimating reliability distribution model parameters, such as parameter λ of the exponential distribution, μ and σ of the normal and lognormal distribution, p of the

161

Reliability Data Analysis and Model Selection

binomial distribution, and α and β of the Weibull distribution. The objective is to find a point estimate and a confidence interval for the parameters of interest. We briefly discussed the MLE concept which relies on the mathematical optimization discussed in Chapter 4 for estimating parameters of probability distribution function. In this approach, an objective function is expressed in terms of a likelihood function. This objective function is a metric expressing the likelihood that the data conform to a hypothesized distribution function. The optimization process finds the parameters of the hypothesized distribution function such that the objective function is maximized. In this section, we expand the application of the MLE to include estimation of parameters of distributions useful to reliability analysis. It is important to appreciate why we need to consider confidence intervals or credible intervals (in case of Bayesian analysis) in this estimation process. In essence, when we have a limited amount of information (e.g., on times to failure), we cannot state our point estimation of the parameters of interest with certainty. Confidence intervals are the formal classical (frequentist) approach to express the uncertainty in our estimations attributed to the limited data set used. The confidence interval is highly influenced by the amount of data available. Of course, other factors, such as diversity and accuracy of the data sources and adequacy of the selected model, can also influence the state of our uncertainty regarding the estimated parameters. When discussing the goodness of fit tests in Chapter 4, we dealt with uncertainty due to the adequacy of the model by using the concept of levels of significance. However, uncertainty due to the diversity and accuracy of the data sources is a more difficult issue to deal with and motivates the use Bayesian approaches as well as appropriate design of data collection and post-processing practices. The method of MLE-based parameter estimation discussed in this section is a more formal and accurate method for determining distribution parameters than the probability plotting and regression methods described previously.

5.4.1

eleMents of Mle using reliability data

Recall that the likelihood functions are objective functions indexes representing the likelihood of observing the reliability data, given the assumed (hypothesized) distribution with pdf f ( t;θ ) with parameter(s), θ . A useful feature of the likelihood function is that it can be used to determine if one distribution is a better fit for a particular set compared to another distribution by solving the likelihood functions for the distributions being compared. The distribution with the higher L value is the better fit. However, this approach may pose problems, because it will be biased toward pdf models with many parameters (due to overfit). The likelihood function is not dependent on the order of each event. Recall Equation 4.5 that defines a likelihood function for a sample of random data di as L (θ data ) ∝ Pr(di θ ) . For complete data, the likelihood function using

∏ all  i

Equation 4.6 would be L (θ x1 ,…, x n ) ∝

n

∏ f ( x θ ) = f ( x θ ) f ( x θ )… f ( x θ ). i

i=1

1

2

n

For censored data, consideration of Equation 4.5 makes it clear that for a left-censored data point, where failure occurred sometime before ti , the probability would be Pr ( T ≤ ti ) = F ( ti ). Similarly, for a right-censored data point for an item which

162

Reliability and Risk Analysis

TABLE 5.1 Likelihood Functions for Different Types of Reliability Data Type of Observation

Likelihood Function

Example Description

Exact lifetimes

Li (θ ti ) = f (ti θ )

Failure time is known.

Left censored

Li (θ ti ) = F ( ti θ )

Component failed before time ti .

Li (θ ti ) = 1 − F ( ti θ ) = R ( ti θ )

Component survived to time ti .

Right censored Interval censored Left truncated Right truncated Interval truncated

(

) (

Li θ ti ) = F (tiRI θ − F tiLI θ f ( ti θ )

Li (θ ti ) =

)

Component failed at time ti where observations are truncated before t L .

R (tL θ ) f ( ti θ )

Li (θ ti ) =

Component failed at time ti where observations are truncated after tU .

F ( tU θ ) f ( ti θ )

Li (θ ti ) =

Component failed between tiLI and tiRI .

Component failed at time ti where observations are truncated before t L and after tU .

F ( tU θ ) − F ( t L θ )

has survived past time ti , Pr ( ti < T ) = 1 − F (ti θ ). For other types of censoring, the contribution to the likelihood function also differs. Table 5.1 gives the constituent parts of the likelihood functions for different types of complete and censored data. Using Table 5.1, we can write the total likelihood function representing multiple complete failure times and right-censored times as

( ) ∏ { f (t θ )

δi

L θ D =c

i

i

( )

1− δ i

× 1 − F ti θ   

}

,

(5.13)

where D stands for all data in the sample, c is a combinatorial constant which quantifies the number of combinations which the observed data could have occurred, and δ i = 1  for complete failure times and δ i = 0 for right-censored times. In MLE process to estimate parameters θ that maximizes Equation 5.13, the constant c will be omitted in the optimization. As we discussed in Chapter 4, a simple way to find the values of θ is to take the partial derivatives of L θ D with respect to each parameter θ and equate it to zero. As such, we will find the same number of equations as unknowns that can produce the point estimates of parameters. Recall that for mathematical simplicity, a common practice is to maximize the log of the objective function under any constraints associated with the parameter(s) θ (e.g., θ > 0). Finding parameters that maximize the logarithmic form of the likelihood function, shown in Equation 5.14 simplifies the computation and maintains the optimal results.

( )

 θˆ = arg max ln[L (θ D)] = arg max  θ θ 

∑ {δ ln  f (t θ ) + (1 − δ ) ln 1 − F (t θ )}. 

i

i

i

i

i

(5.14)

163

Reliability Data Analysis and Model Selection

A convenient approach to find the confidence intervals of θ in Equation 5.13 is to use the Fisher information matrix approach. The approach has many uses besides finding the intervals of the parameters of reliability pdf, including application to the Jeffery’s non-informative priors used as a non-informative prior distribution in the Bayesian parameters estimation. The Fisher information matrix is obtained from using the log-likelihood function Λ (θ D ) = ln  L θ D  and replacing the parameters by their point estimates  θˆ .   For a single-parameter distribution, the Fisher information of the parameter θ is obtained from:

(

)

 ∂2 Λ (θ D )  I θˆ = −   . 2 θ =θˆ  ∂θ

()

(5.15)

The observed Fisher information matrix is the negative of the second derivative of the of the log-likelihood function (this second derivative is also called the Hessian matrix). So, for a pdf model with a vector of p parameters, θ , using the log-likelihood Λ (θ D ) for all data D discussed earlier, the observed Fisher information is expressed by a p × p symmetric matrix in the form of

(

)

−

(

)

−

(

)

 2  −∂ Λ θ D  ∂θ12   ∂2 Λ θ D  − ˆ I θ = ∂θ 2 ∂θ1    2  ∂ Λ θ D  −  ∂θ p ∂θ1 

()

(

)



−

(

)



−

∂2 Λ θ D ∂θ1 ∂θ 2 ∂2 Λ θ D ∂θ

2 2

 −

(

∂ Λ θ D 2

∂θ p ∂θ 2

)

)

(

)

∂θ1 ∂θ p ∂2 Λ θ D ∂θ 2 ∂θ p



 −

(

∂ Λ θ D 2



(

∂2 Λ θ D

∂θ

2 p

)

      (5.16)       θi =θi

When the inverse of the I (θ ) matrix is evaluated at θ = θˆ , we find the observed Fisher information matrix in Equation 5.16 produces the variance-covariance matrix in Equation 5.17:

()

()

Var θˆ =  I θˆ   

−1

 Var (θ ) 1   Cov (θ ,θ ) 2 1 =    θ Cov ( p ,θ1 )  

Cov (θ1 ,θ 2 )



Var (θ 2 )



 Cov (θ p ,θ 2 )

 

Cov (θ1 ,θ p )   Cov (θ 2 ,θ p )  .    Var (θ p )  

(5.17)

The variance-covariance matrix provides information about the dependencies between parameters (in terms of their covariances), and its diagonal also shows the

164

Reliability and Risk Analysis

variance of each parameter. If we use this variance-covariance matrix with a large number of samples, the asymptotic normal property can be used to estimate confidence intervals. By connecting this to percentiles of the normal distribution, 100 γ % approximate confidence intervals are generated. If the range of a single parameter θ is unbounded ( −∞,∞ ) the approximate two-sided lower confidence limit θ l and upper confidence limit θ u are given by 1+γ  θ l ≈ θˆ − Φ−1  Var θˆ ,  2 

()

(5.18)

()

(5.19)

and 1+γ  θ u ≈ θˆ + Φ−1  Var θˆ .  2 

If the range of θ is ( 0,∞ ) the corresponding approximate two-sided confidence limits are

()

θ l ≈ θˆ ⋅ e

 −1  1+ γ  ˆ   Φ  2  Var θ   −   θˆ    

,

(5.20)

.

(5.21)

and

()

θ u ≈ θˆ ⋅ e

 −1  1+ γ  ˆ   Φ  2  Var θ      θˆ    

If the range of θ is ( 0,1) the similar approximate two-sided confidence limits are −1

 −1  1+γ  ˆ    Φ  2  Var (θ )     ˆ ˆ   θ (1−θ )     θ l ≈ θˆ ⋅ θˆ + 1 − θˆ e  ,  

( )

(5.22)

and −1

 −1  1+γ  ˆ    Φ  2  Var (θ )    −   θˆ(1−θˆ )     . θ u ≈ θˆ ⋅ θˆ + 1 − θˆ e   

( )

(5.23)

Reliability Data Analysis and Model Selection

165

The advantage of this approximate method is that it can be calculated for all distributions and is easy to determine. The disadvantage is that the assumption of a normal distribution is asymptotic and so sufficient data is required for the confidence interval estimate to be accurate. The number of samples needed for an accurate estimation changes from distribution to distribution. It also produces symmetrical confidence intervals which may be approximate. As discussed in Chapter 4, understanding how to interpret the confidence intervals found from the classical statistical methods described in this section is critical. For a sample of data, it is possible to say that the probability that the associated confidence interval contains the true value of a parameter of interest is (1 − γ ). Further, one can also interpret the interval such that if we obtained many similar samples to the one used to estimate the confidence interval, 100 (1 − γ ) % of the associated confidence intervals will contain the true value of the parameter. Therefore, the confidence level in the classical statistical estimation refers to the confidence over the interval not the parameter of interest.

5.4.2

exPonential distribution Mle Point estiMation

The MLE point estimator for the exponential distribution parameter is r λˆ = , TTT

(5.24)

where r is the number of failures observed and TTT is the total time on test. Correspondingly, the MLE for the MTTF is given by TTT  M TTF = . r For complete data, TTT = depending on test design.

(5.25)

∑ t . For right-censored data, the total time on test differs i

5.4.2.1 Type I Life Test with Replacement Suppose n components are placed on reliability test with replacement (i.e., monitored and replaced when failed), and the test or field data collection is terminated after a specified time tend . The total accumulated test time from both failed and rightcensored components, TTT, (in hours or cycles), is given by TTT = n tend .

(5.26)

Equation 5.26 shows that at each time instant from the beginning of the test up to time tend , exactly n components have been on test (i.e., under observation). Accordingly, if r failures have been observed up to time tend for a component with times to failure following the exponential distribution, the maximum likelihood point estimate of the failure rate of the component can be found by using Equation 5.25.

166

Reliability and Risk Analysis

The number of items tested during the test, n′, is n′ = n + r.

(5.27)

5.4.2.2 Type I Life Test without Replacement Suppose n components are placed on test without replacement, and the test (or field observation) is terminated after a specified time tend during which r failures have occurred. The total time on test, for the failed and survived components is r

TTT =

∑t + ( n − r ) t i

end

,

(5.28)

i=1

r

where

∑t represents the accumulated time on test of the r failed components (note i

i =1

that r is a random variable here), and ( n − r ) tend is the accumulated time on test of the surviving (right-censored) components at the end of the test. Since no replacement has taken place, the total number of components tested in the test is n′ = n. Note that if items are right censored before the end of the test, their contributions to the total time on test are not reflected in Equation 5.28. To account for those events, the sum of the right-censored times for such items should be added to Equation 5.28. 5.4.2.3 Type II Life Test with Replacement Consider a situation in which n components are being tested (or field observed) with replacement, and a component is replaced with an identical component as soon as it fails (except for the last failure). If the test is terminated after a time tr when the rth failure has occurred (i.e., r is specified but tr is random), then the total time on test associated with both the failed and right-censored components is given by TTT = ntr .

(5.29)

Note that tr, unlike tend , is an r.v. The total number of items tested, n′, is n′ = n + r − 1, 

(5.30)

where ( r − 1) is the total number of failed and replaced components. All failed components are replaced except the last one, because the test is terminated when the last component fails (i.e., the rth failure has been observed). 5.4.2.4 Type II Life Test without Replacement Consider another situation when n components are being tested without replacement, that is, when a failure occurs, the failed component is not replaced by a new one. The

167

Reliability Data Analysis and Model Selection

test is terminated at time tr when the rth failure has occurred (i.e., r is specified but tr is random). The total time on test of both failed and right-censored components at the end of the test is obtained from r

TTT =

∑t + ( n − r )t , i

r

(5.31)

i =1

r

where

∑t is the accumulated time contribution from the failed components and i

i =1

( n − r ) tr   is the accumulated time contribution from the surviving components. It should also be noted that the total number of items tested is n′ = n, 

(5.32)

since no components are being replaced. Note that if items are right censored before the end of the test, their contributions to the total time on test are not reflected in Equation 5.31. To account for those events, the sum of the right-censored times for such items should be added to Equation 5.31. Example 5.7 Ten light bulbs are placed under life test. The test is terminated at tend = 850 hours. Eight components fail before 850 hours have elapsed. The failure times obtained are 183, 318, 412, 432, 553, 680, 689, and 748. Determine the total accumulated component hours and estimate of the failure rate and MTTF for these situations: a. b. c. d.

The components are replaced when they fail. The components are not replaced when they fail. Repeat a, assuming the test is terminated when the eighth component fails. Repeat b, assuming the test is terminated when the eighth component fails.

Solution: a. For a Type I test with replacement, Using Equation 5.26, TTT = 10 ⋅850 = 8,500 component hours.

λˆ =

8 1  TTF = = 1,062.5 hours. = 9.4 × 10 −4 / hour and M 8,500 λˆ

b. For a Type I test without replacement r

∑t = 4,015 and (n − r )t i

i=1

end

= (10 − 8 ) 850 = 1,700.

168

Reliability and Risk Analysis Thus, TTT = 4,015 + 1,700 = 5,715 component hours. 8  = 1.4 × 10 −3 / hour and M TTF = 714.4 hours. 5715

λˆ =

c. For a Type II test with replacement Here, tr is the time to the eighth failure, which is 748. TTT = 10(748) = 7,480 component hours. 8  = 1.1 × 10 −3 / hour and  M TTF = 935 hours. 7,480

λˆ =

d. For a Type II test without replacement r

∑T = 4,015 and (n − r )t i

end

= (10 − 8 ) 748 = 1,496.

i=1

Thus, TTT = 4,015 + 1,496 = 5,511 component hours.

λˆ =

8  = 1.5 × 10 −3 / hour and M TTF = 688.8 hours. 5,511

A simple comparison of the results shows that although the same set of data is used, the effect of the type of the test and of the replacement of the failed items could be significant.

5.4.3

exPonential distribution interval estiMation

In Section 5.4.2, we discussed the MLE approach to find the point estimate of parameter λ (i.e., failure rate) of the exponential distribution. This point estimator is λˆ = r / T , where r is the number of failures observed and T is the total observed time (often, the total time on test, TTT). Epstein (1960) has shown that if the time 2rλ to failure is exponentially distributed with parameter λ, the quantity = 2λ ⋅ TTT λˆ has the χ 2 distribution with 2r   degrees of freedom for Type II censored data (failureterminated test). Based on this information, one can construct the corresponding confidence intervals. Because uncensored data can be considered as a particular case of Type II right-censored data (when r = n), the same procedure applies to the complete (uncensored) sample. Using the distribution of 2rλ / λˆ at the (1 − γ ) confidence level, one can write   2rλ Pr  χ (2γ /2) [ 2r ] ≤ ≤ χ 2 γ  [ 2r ] = 1 − γ .  1−  λˆ  2 

(5.33)

In Equation 5.33, the χ 2 is evaluated at the 2r degrees of freedom for a significance level of γ /2 for the lower limit, and 1 − γ / 2 significance level for the upper limit. The

169

Reliability Data Analysis and Model Selection

values of χ 2 are available from Table A.3. By rearranging Equation 5.33 and using λˆ = r / TTT, the two-sided confidence interval for the true value of λ can be obtained:  χ 2γ  [ 2r ] χ 2 γ  [ 2r ]  1−   2    2  Pr  ≤λ≤ = 1 − γ .  2TTT 2TTT     

(5.34)

The corresponding upper confidence limit (the one-sided confidence interval) is  χ (21−γ ) [ 2r ]  Pr  0 ≤ λ ≤  = 1 − γ .  2TTT  

(5.35)

Accordingly, confidence intervals for MTTF and R(t) at a time t = tend can also be obtained as one-sided and two-sided confidence intervals from Equations 5.34 and 5.35. The results are summarized in Table 5.2. TABLE 5.2 100 (1−γ)% Confidence Limits on λ, MTTF, and R(t) One-Sided Confidence Limits Parameter

Lower Limit

Upper Limit

Two-Sided Confidence Limits Lower Limit

Upper Limit

Type I (Time Terminated Test for Complete Data) λ

0

χ (21−γ ) [ 2r + 2 ]

χ 2γ  [ 2r ]  2 

2TTT MTTF

2TTT χ (21−γ ) [ 2r + 2 ]

R(t0) e

  χ2 (1− γ ) [ 2 r + 2] −  tend  2TTT  

χ 2

γ 1−  2 

2TTT

∞

χ 2

2TTT [ 2r + 2]

γ 1−  2 

1

e

 χ2 [ 2r + 2]    1− γ     2  tend  − 2TTT      

[ 2r + 2]

2TTT 2TTT χ 2γ  [ 2r ]  2 

e

  χ 2 [ 2r ]    γ     2 tend  −   2TTT    

Type II (Failure Terminated Test for Complete Data) λ

0

χ (21−γ ) [ 2r ]

χ 2γ  [ 2r ]  2 

2TTT MTTF

2TTT χ (21−γ ) [ 2r ]

R(t0) e

  χ2 (1− γ ) [ 2r ] −  tend  2TTT  

∞

χ 2

γ 1−  2 

2TTT 2TTT χ 2 γ  [ 2r ] 1−  2 

1 e

 χ2 [ 2r ]    1− α     2  tend  −   2TTT    

[ 2r ]

2TTT 2TTT χ 2γ  [ 2r ]  2 

e

  χ 2 [ 2r ]    γ     2 − tend    2TTT    

170

Reliability and Risk Analysis

As opposed to Type II censored data, the corresponding exact confidence limits for Type I censored data are not available. However, the approximate two-sided confidence interval for failure rate λ for Type I (time-terminated test) data usually is constructed as  2 χ 2 γ  [ 2r + 2 ]   1−    χ (γ / 2) [ 2r ] 2 ≤λ≤ Pr   = 1 − γ .  2TTT 2TTT    

(5.36)

The χ 2 in Equation 5.36 is evaluated at the 2r degrees of freedom for a significance level of γ /2 for the lower limit, and at the (2r + 2)  degrees of freedom and (1 − γ / 2) significance level for the upper limit. The respective upper confidence limit (a one-sided confidence interval) is given by  χ (21−γ ) [ 2r + 2 ]  Pr  0 ≤ λ ≤  = 1 − γ .  2TTT  

(5.37)

If no failure is observed during a test, the formal estimation gives λˆ = 0, or MTTF = ∞. This cannot realistically be true, since we may have had a small or limited observations with no failures, or a failure may be about to occur before the test ended. Had the test been continued, eventually a failure would have been observed. Therefore, an upper confidence estimate for λ can be obtained for r = 0. However, the lower confidence limit cannot be obtained with r = 0. It is possible to relax this limitation by conservatively assuming that a single failure occurs exactly at the end of the observation period. Then r = 1 can be used to evaluate the lower limit for the two-sided confidence interval. This conservative modification, although sometimes used to allow a complete statistical analysis, lacks a firm statistical basis. This limitation is relaxed in the Bayesian estimations of these parameters, as will be discussed later in this chapter. Example 5.8 Twenty-five items are placed on a reliability test that lasts 500 hours. In this test, eight failures occur at 75, 115, 192, 258, 312, 389, 410, and 496 hours. The failed items are replaced. Assuming a constant failure rate, find λˆ , the one-sided and two-sided confidence intervals for λ, and MTTF at the 90% confidence level; and one-sided and two-sided 90% confidence intervals for reliability at t = 1,000 hours.

Solution: This is a Type I test. The accumulated time TTT is obtained from Equation 5.26 TTT = 25 ⋅ 500 = 12,500 hours.

171

Reliability Data Analysis and Model Selection The point estimate of failure rate is

λˆ = 8/12,500 = 6.4 × 10 −4 / hour. One-sided confidence interval for λ is

0≤λ≤

χ 2 [2 ⋅ 8 + 2] . 2 ⋅ 12,500

Finding the inverse of the χ 2 distribution yields: 2 χ 0.9 (18) = 25.99,  0 ≤ λ ≤ 1.04 × 10 −3 / hour.

Two-sided confidence interval for λ is 2 2 χ 0.05 [2 ⋅ 8] ≤ λ ≤ χ 0.95 [2 ⋅ 8 + 2] . 2 ⋅ 12,500 2 ⋅ 12,500

Finding the inverse of the χ 2 distribution yields: 2 2 χ 0.05 (16) = 7.96 and χ 0.95 (18) = 28.87.

Thus, 3.18 × 10 −4 / hour ≤ λ ≤ 1.15 × 10 −3 / hour. One-sided 90% confidence interval for R(1,000) is e

(

)

− 1.04 ×10 −3 (1,000 )

≤ R (1,000 ) ≤ 1,

or 0.35 ≤ R (1,000 ) ≤ 1. Two-sided 90% confidence interval for R(t) is e

(

)

− 1.15×10 −3 (1,000 )

≤ R (1,000 ) ≤ e

(

)

− 3.18 ×10 −3 (1,000 )

or 0.32 ≤ R (1,000 ) ≤ 0.73.

,

172

Reliability and Risk Analysis

5.4.4

norMal distribution

The normal distribution was introduced in Chapter 4 in the context of complete data. The key equations are presented here again for completeness of the chapter. For the normal distribution, the confidence interval for µ   with σ   known is:     x−µ Pr  − z γ  ≤ ≤ z  γ   = 1 − γ ,  σ  1−   1−  2 2     n where z

γ  1−  2

(5.38)

is the 100 (1 – γ / 2 ) % of the standard normal distribution (which can be

obtained from Table A.1) and x is the sample mean. In the case when σ is unknown and is estimated as s, e.g., using Equation 4.3, the respective confidence interval on µ is given by:  s s  Pr  µˆ − t γ  [ n − 1] < µ < µˆ + t γ  [ n − 1] = 1 − γ ,  n n    2   2 

(5.39)

where t(γ / 2)   is the percent of the one-sided student’s t-distribution with (n – 1) degrees of freedom. Values of tγ for different numbers of degrees of freedom are given in Table A.2. Similarly, confidence intervals for σ 2 for a normal distribution can be obtained as

( n − 1)σˆ 2

χ 2

γ  1−  2

where χ 2

γ  1−  2

[ n − 1]

n /10, Equation 5.36 is not a good approximation.

5.5 CLASSICAL NONPARAMETRIC DISTRIBUTION ESTIMATION We have already established that any reliability measure or index can be expressed in terms of the time to failure cdf or reliability function. Thus, the problem of estimating these functions is important. We will now turn to how to achieve this using nonparametric methods. The commonly used estimate of the cdf is the empirical (or sample) distribution function (edf) introduced for uncensored data in Chapter 4 (see Equation 4.28) in the context of the K-S test. In this section, we consider some other nonparametric point and confidence estimation procedures applicable for censored data.

5.5.1

confidence intervals for cdf and reliability function for coMPlete and censored data

Constructing an edf requires a complete sample, but an edf can also be constructed for the right-censored samples for the failure times which are less than the last time to failure observed (t < tr). The edf is a random function, since it depends on the sample items. For any point, t, the edf, Sn ( t ), is the fraction of sample items that failed before t.

180

Reliability and Risk Analysis

The edf is, in a sense, the estimate of the probability, p = F ( t ), in a binomial trial. It is possible to show that the MLE of the binomial parameter p coincides with the sample cdf Fn ( t ) = i / n (Equation 4.28) and that Fn ( t ) is a consistent estimator of the cdf, F(t). Using the relationship between reliability and the edf, one can obtain an estimate of the reliability function. This estimate, called the empirical (or sample) reliability function, is  1,  i  Rn ( t ) =  1 − , n  0, 

0 < t < t1; ti ≤ t < ti +1 , i = 1,2,…, n − 1;

(5.62)

tn ≤ t < ∞;

where t1 <  < tn are the ordered failure data. The mean number of failures observed during time t is E ( r ) = np = nF ( t ), and so the mean value of the proportion of items failed before t is E ( r / n ) = p = F ( t ) . The variance of this fraction is given by p (1 − p ) F ( t )[1 − F ( t )] r = .  Var   =  n n n

(5.63)

For practical problems, Equation 5.63 is used by replacing F(t) with Fn(t). As the sample size, n, increases, the binomial distribution can be approximated by a normal distribution with the same mean and variance [i.e., μ = np, σ 2 = np(1−p)]. This approximation provides reasonable results if both np and n(1−p) are ≥5. Using this approximation, the 100 (1 – γ /2 ) % confidence interval for the unknown cdf, F(t), at any point t can be constructed as:

Fn ( t ) − z

γ  1−  2

where z

γ  1−  2

Fn ( t ) [1 − Fn ( t )] Fn ( t ) [1 − Fn ( t )] ≤ F ( t ) ≤ Fn ( t ) + z γ  , 1−   n n 2

(5.64)

γ is the inverse of the probability of level  1 −  of the standard normal  2

distribution. The corresponding estimate for the reliability (survivor) function is Rn = 1 − Fn ( t ) . Example 5.13 Using the data from Example 5.4, find the nonparametric point estimate and the 95% confidence interval for the cdf, F(t) at t = 350 hours.

Solution: Using Equation 5.59 the point estimate for F(350) = Fn(350) = 5/10.

181

Reliability Data Analysis and Model Selection The respective approximate 95% confidence interval based on Equation 5.64 is 0.5 − 1.96

0.5 (1 − 0.5) 0.5 (1 − 0.5) ≤ F ( 350 ) ≤ 0.5 + 1.96 . 10 10

Therefore, Pr(0.190 < F ( 350 ) < 0.8099) = 0.95.

Using complete or right-censored reliability data from an unknown cdf, one can also obtain the strict confidence intervals for the unknown cdf, F(t). This can be done using the same Clopper–Pearson procedure for constructing the approximate confidence intervals for the binomial parameter p, using Equations 5.60 and 5.61. However, these limits can also be expressed in more compact form in terms of the incomplete beta function as follows. The lower confidence limit, Fl ( t ), at the point t where Sn ( t ) = i / n (r = 0,1,…, n), is the largest value of p that satisfies the following inequality I p ( r , n − r + 1) ≤

γ , 2

(5.65)

and the upper confidence limit, Fu ( t ), at the same point is the smallest p satisfying the inequality I1− p ( n − r , r + 1) ≤

γ , 2

where

(5.66)

p

B ( p; α , β ) β −1 I p (α , β ) =  with B (⋅)  defined as B ( p; α , β ) = uα −1 (1 − u ) du. B (1; α , β )

∫ 0

is known as I p (α , β ) the regularized incomplete beta function which represents the cdf of the beta distribution. The regularized incomplete beta function is difficult to tabulate; however, its numerical approximations are available online. Example 5.14 For the data from Example 5.13, find the 95% confidence interval for the cdf, F(t) at t = 350 hours, using Equations 5.65 and 5.66.

Solution: Using Equation 5.65, the lower confidence limit is found from I p ( 5,10 − 5 + 1) ≤ 0.025. Solving for the largest p, the inequality above yields a value for pl = 0.1871  and, using Equation 5.66, the inequality for the upper confidence limit below leads to pu = 0.8131.

182

Reliability and Risk Analysis

183

Reliability Data Analysis and Model Selection

Solution: When r = 0, Equation 5.67 can be written as I 0.86 ( n − 1,1) ≤ 0.1. The resulting

value is n = 16. For r = 1 Equation 5.67 can be written as I 0.86 ( n − 1,2 ) ≤ 0.1 which yields the minimum number of items needed for test as n = 26.

5.5.2

confidence intervals of reliability function for censored data

The point and confidence estimates considered so far do not apply when censored items occur during the test. For such samples, the Kaplan-Meier or product-limit estimate, which is the MLE of the cdf (unreliability function), can be used. Suppose we have a sample of n times to failure, among which there are only k distinct failure times. Consider the ordered times of failure as  t1 ≤  ≤ t k , and let t0 = 0. Let nj be the number of items under observation just before tj. Assume that the time to failure pdf is continuous and there is only one failure at a time. Then, nj + 1 = (nj + 1). Under these conditions, the Kaplan-Meier estimate is given by    Fn ( t ) = 1 − Rn ( t ) =  1 −   

0 ≤ t < t1 ,

0, i

∏ j =1

nj − 1 , nj

ti ≤ t < ti +1 , i = 1,…, m − 1,

(5.68)

t ≥ tm ,

1,

where integer m = k, if k < n, and m = n, if k = n. It is possible to show that for complete data samples, the Kaplan-Meier estimate coincides with the edf given by Equation 4.28. In the general case when there are multiple failures, d j at time t j or during an interval, the Kaplan-Meier estimate is given by    Fn ( t ) = 1 − Rn ( t ) =  1 −   

0 ≤ t < t1 ,

0, i

∏ j =1

nj − d j , nj

ti ≤ t < ti +1 , i = 1,…,m − 1,

(5.69)

t ≥ tm ,

1,

For estimation of variance of Sn (or Rn), Greenwood’s formula (Lawless, 2011) is used: Var  Fˆn ( t )  = Var  Rˆ n ( t )  =

∑ n (n − d ) dj

j :t j